CN101835045A - Hi-fidelity remote sensing image compression and resolution ratio enhancement joint treatment method - Google Patents

Abstract

The invention relates to a hi-fidelity remote sensing image compression and resolution ratio enhancement joint treatment method, relating to a remote sensing image treatment method and solving the problem that the conventional remote sensing image compression method tends to abandon high resolution information without distinction to cause the reduction of recovering image resolution ratio. The method comprises steps of: 1, inputting an image f (x, y) to be compressed; 2, preprocessing the image; 3, carrying out discrete wavelet transform on processed data; 4, carrying out information detection on the result after the wavelet transform; 5, quantizing the result after the wavelet transform and carrying out subband weighting simultaneously; 6, carrying out entropy coding on the quantization result of the step 5; 7 carrying out code stream interception on the entropy coding result and carrying out channel weighting simultaneously; and 8, obtaining compression code stream. Aiming at the actual application requirement of the existing remote sensing image, the invention has important value in application of compressed images in aspects of edge detection, target identification and the like.

Description

Combined processing method for high-fidelity remote sensing image compression and resolution enhancement

Technical Field

The invention relates to a remote sensing image processing method, in particular to a remote sensing image processing method combining compression and resolution maintenance.

Background

With the development of remote sensing technology, remote sensing images are increasingly widely applied in the aspects of earth resource management, environmental monitoring, military reconnaissance and the like. However, the problem that the data volume is too large needs to be solved preferentially in the application. The mass data brings great difficulties to transmission, storage and processing, so that effective data compression on the remote sensing image is particularly urgent and important.

The compression of the remote sensing image has commonality compared with the compression of a common digital image, and the data redundancy is removed by a certain method, but the compression of the remote sensing image has particularity. Most remote sensing images have the characteristics of high resolution, large information amount and low redundancy, so the compression effect obtained by the common compression method is not ideal. Moreover, the application background of the remote sensing image is special, and the compression algorithm which only uses the objective quality such as signal-to-noise ratio and the like as the evaluation criterion and abandons a large amount of high-frequency components and reserves low-frequency information is not suitable. Because the high-frequency components of the remote sensing image are likely to contain interesting information such as small objects, textures and the like, the information is saved as much as possible in the compression of the remote sensing image. Therefore, it is necessary to research a suitable compression method according to the characteristics of the remote sensing image and the special application background thereof.

The latest international compression standard of still images, JPEG2000, provides a new solution for remote sensing image compression, and the excellent characteristics of rate distortion performance, details, small target holding capacity and the like are very suitable for remote sensing image compression. However, even JPEG2000 with high fidelity performance has certain limitations for remote sensing image compression. The JPEG2000 measures the importance of information based on objective quality, and for high-frequency information in the same resolution, it is considered that the importance of any information in the resolution is lower than that of information in the previous resolution, and a large amount of high-frequency components tend to be discarded in the compression process, so that a large amount of high-resolution information is lost, or even all high-resolution information is lost, which not only results in the reduction of the resolution of the restored image, but also some detailed information is easily damaged or even lost, which is often of interest to users and is required to have higher fidelity. In addition, the remote sensing image has many special applications such as edge detection, object recognition and the like besides serving a human eye visual system. The indiscriminate loss of large amounts of high-resolution information tends to result in a reduction in image resolution, which is detrimental to both the visual effect on the human eye and for later applications. In order to compensate for the reduced image resolution caused by compression, corresponding resolution enhancement processing is often performed. Most resolution enhancement techniques are currently used as only one type of post-processing. However, in the compression process, once the resolution information is lost, the resolution enhancement technology cannot be restored, and the resolution enhancement technology as the post-processing only uses the correlation between pixels to predict and compensate the lost information through other information, so that the effect of approximate resolution enhancement is achieved subjectively by human eyes or computer processing, but the information amount of the restored image is not increased in fact. Only by embedding the resolution enhancement technology in the compression process, the resolution reduction of the image caused by the compression can be effectively prevented, and the resolution of the restored image can be really enhanced. Especially for the remote sensing image compression with rich high-frequency information and special application background, the combined processing of compression and resolution enhancement is considered, and important high-resolution information is preferentially reserved as far as possible on the premise of ensuring the overall quality of the restored image, so that higher subjective and objective quality is obtained, and the method is favorable for later special application. However, the research on the combination of compression and resolution enhancement is still in the beginning and still needs to be further explored.

Disclosure of Invention

The method aims to solve the problem that the resolution of a recovered image is reduced because a common remote sensing image compression method tends to discard high-resolution information indiscriminately. And a combined processing method of high-fidelity remote sensing image compression and resolution enhancement is provided.

The combined processing method for compressing and enhancing the resolution of the high-fidelity remote sensing image comprises the following steps:

the method comprises the following steps: inputting an image f (x, y) to be compressed;

step two: preprocessing the image;

step three: performing discrete wavelet transform on the processed data;

step four: carrying out information detection on the result of the wavelet transform;

step five: carrying out quantization processing on the wavelet transform result and simultaneously carrying out sub-band weighting;

step six: entropy coding is carried out on the quantization result of the step five;

step seven: carrying out code stream truncation on the entropy coding result, and simultaneously carrying out channel weighting;

step eight: and obtaining a compressed code stream.

Aiming at the actual application requirements of the existing remote sensing images, the invention provides a joint processing method for high-fidelity remote sensing image compression and resolution enhancement based on JPEG2000 standard, and the method is suitable for remote sensing images and general images. In conclusion, the combined processing method for high-fidelity remote sensing image compression and resolution enhancement has profound practical significance and application value for development of remote sensing technology and application of remote sensing images, has certain foresight, provides a new direction for remote sensing image compression, and has important value for application of compressed images such as edge detection, target identification and the like.

In the joint processing method, in order to enhance the resolution of an image while compressing, relatively important information needs to be detected before compression, that is, information needs to be embedded in the compression process for detection, and the important information mainly includes target and edge information thereof.

The information detection method adopted in the fourth step is a multi-resolution edge detection algorithm. According to the multiresolution decomposition and time-frequency localization analysis characteristics of wavelet transformation, the method for detecting the image edge comprises the following steps: self-adaptively setting double thresholds TH of each sub-band_b1And TH_b2The method comprises the steps of detecting multi-resolution edge information by a double-threshold method, matching and denoising the edge information with different resolutions, and extracting the edge information of an interested target.

The main processes for implementing compression by the JPEG2000 standard are known: JPEG2000 mainly implements rate control by quantization and PCRD truncation, and theoretically, information loss is mainly concentrated in these two parts. The saving of the multi-resolution edge information also needs to be done separately for both parts.

And step five, quantizing the wavelet transform result, weighting the sub-bands simultaneously, not adopting a non-uniform quantization method, still using the original quantization mode of JPEG2000, and only integrally adjusting the quantization step size of each sub-band. HH with highest resolution₁The sub-band is used as a reference, the quantization step size is set to be 1, the relative proportion relation between the rest sub-bands and the sub-band is kept unchanged, and the adjustment is carried out according to the proportion. After adjustment, the quantization step size of all sub-bands is less than or equal to 1, and quantization will not cause information loss. The reason for this is mainly based on the following two points:

(1) unlike other compression methods, the rate control of JPEG2000 consists of two parts, quantization and PCRD truncation. Quantization is only implemented once, which is a rough rate control, and the resulting rate is necessarily different from the target rate. Finer and final rate control is achieved by the PCRD by selecting the coding pass. It can also be known from the basic principle of JPEG2000 that, because JPEG2000 adopts fragmented bit plane coding and embedded coding, the PCRD truncation algorithm selects a code stream better than the quantization method, that is, JPEG2000 does not need to rely on quantization alone to realize code rate control. Indeed, to achieve higher restored image quality, JPEG2000 itself is more prone to prefer codestreams by PCRD truncation.

(2) The purpose of the multi-resolution edge information preservation is to prevent relatively important information from being lost as much as possible, and all information can be preserved during quantization as long as the quantization step size is not greater than 1. The cost is that the non-important information is also saved together, i.e. the quantization not only does not reduce the amount of data, but also adds extra redundant data, which will increase the burden of encoding to some extent. However, this method is superior to the coding burden imposed by the non-uniform quantization method; the added redundant code stream can be cut off through a PCRD algorithm, and extra bit streams cannot be added in the target code stream, so that the image quality cannot be greatly influenced; moreover, the method does not need to change a decoder, and any decoder supporting JPEG2000 code streams can normally decompress the code streams.

This ensures that the detected multiresolution edge information is not lost by quantization.

Seventhly, code stream truncation is carried out on the entropy coding result, and channel weighting is carried out at the same time;

the PCRD algorithm of JPEG2000 selects the channel which contributes most to the restored image from all channels of all code blocks on the premise of meeting the requirement of a given target code rate, and abandons the channel which contributes relatively less to the restored image, thereby minimizing distortion as far as possible. The rate-distortion slope is used to measure the amount of contribution of each channel to the recovered image. That is, the PCRD algorithm preferentially selects and retains the channel with the larger rate-distortion slope, and discards the channel with the smaller rate-distortion slope. Therefore, to prevent the multi-resolution edge information from being truncated and lost, the channel where the multi-resolution edge information is located can be weighted to make the rate-distortion slope value larger than that of other channels, so that the multi-resolution edge information can be automatically selected and retained by the PCRD algorithm when the layer two is truncated.

Theoretically, most or even all of the multi-resolution edge information can be saved by weighting the channel where the multi-resolution edge information is located by a large margin, so that the target edge of the restored image is nearly lossless or even lossless. However, on the premise that the target code rate is fixed, the storage of the multi-resolution edge information increases the code stream occupied by the high-frequency information in the target code rate, and the information of the corresponding other frequency segments is discarded and lost. Since the information of these frequency bands has a large global contribution to the restored image relative to the high-frequency information, and the multi-resolution edge information has a significant contribution to only a local part of the restored image, or even only the edge of the target, this may result in a reduction in the global quality of the restored image, while when the information discarded by the replacement is low-frequency information, especially the channel of the LL subband, this may also result in a reduction in the contrast of the entire image including the target, which is rather irrevocable. Therefore, compromise storage is needed between multi-resolution information and other frequency band information, and the combined processing of compression and resolution enhancement is realized on the premise of ensuring the overall quality of the restored image.

In order to check the effectiveness of the high-fidelity remote sensing image compression and resolution enhancement combined processing algorithm provided by the invention, a JPEG2000Part 1 recommendation algorithm and a Jasper algorithm recommended by Part 5 are used as comparison objects, the comprehensive effect of the algorithm in the aspects of ensuring the image quality to be restored and enhancing the edge resolution is evaluated by selecting the global and local peak signal-to-noise ratio, the subjective quality, the edge detection effect and the like of the restored image, and the calculation complexity of the algorithm is evaluated by providing the compression time of the three algorithms.

The test image selected for simulation is a satellite-borne optical remote sensing image with the size of 1024 × 1024 × 8 bits (see fig. 1, the image is too large and is reduced to 33% of the original size for display), the experimental parameters respectively select 9/7 wavelet bases, the 4-layer wavelet decomposition is performed, the code block size is 64 × 64, and the target code rate (unit bpp) respectively is: 1.0, 0.5, 0.30, 0.25, 0.125, corresponding compression ratios are: 8. 16, 27, 32 and 64 times. In order to examine the effect of the joint processing algorithm provided by the invention on the high-fidelity compression of the target and the enhancement effect of the high-fidelity compression on the resolution ratio of the target, two local regions with rich edge information are selected, and are tested in fig. 2 and fig. 3 (the display size of each local region is 75% of the original size).

The PSNR comparison of the entire restored image and its two local regions is shown in table 3-1, where the "-" number indicates the PSNR degradation of the algorithm of the present invention compared to Jasper and Part 1 algorithms. As can be seen from the data in the table, for the edge information, namely the important and richer local area, the PSNR of the algorithm is greatly improved compared with the Jasper algorithm and the Part 1 algorithm. For example, when the compression is 16 times, for the PSNR of the first local area, the algorithm is improved by 2.163dB compared with a Jasper algorithm, and is improved by 2.632dB compared with a Part 1 algorithm; for the PSNR of the second local area, compared with a Jasper algorithm, the algorithm of the invention is improved by 3.014dB, and compared with a Part 1 algorithm, the algorithm is improved by 3.648 dB. This is because, the algorithm of the present invention preferentially protects the multiresolution edge information of the target during the compression process, so that the edge information is selected earlier than the background and other information and enters the target code stream, therefore, for the region with rich edge information in the restored image, the algorithm of the present invention has higher fidelity than the Jasper algorithm and Part 1 algorithm, and because the edge resolution is effectively stored, the edge resolution of the restored image is enhanced compared with both the Jasper algorithm and Part 1 algorithm. The reason why the PSNR of the second local area is improved more significantly than that of the first local area is that the edge information of the second local area is richer than that of the first local area, which results in more multi-resolution edge information being stored in the second local area, and the quality of the restored image is naturally improved more significantly. Of course, the global quality of the restored image is slightly degraded since the high-resolution edge information is more preserved and accordingly other information is lost by replacement. However, the ratio of the interested edge information to the image information is small, so that the storage of the edge information has little influence on the quality of the whole image. In terms of global quality, under different compression ratios, the reduction amplitude of the algorithm is only about 0.5dB compared with that of a Jasper algorithm, and is about 0.2dB compared with that of a Part 1 algorithm.

TABLE 3-1 PSNR comparison (Unit: dB) of global and local regions of restored image

Besides objective quality evaluation, the invention also provides subjective quality evaluation. Because the original image is too large, the subjective effect is not easily seen after the original image is reduced, and because of space limitation, the invention only provides the recovery images of three different algorithms when the second local area is compressed by 16 times (as in fig. 4 to 6) and 32 times (as in fig. 7 to 9).

From the subjective effect, the target information recovery quality of the airplane, the house and the like of fig. 6 and 9 is obviously better than that of fig. 4, 5, 7 and 8, the edge contour is clearer, the edge resolution is higher, and the same effect is achieved under other compression ratios. Therefore, the recovery quality of the algorithm to the target is higher in fidelity than that of the Jasper algorithm and the Part 1 algorithm, and the target edge resolution is also higher than that of the Jasper algorithm and the Part 1 algorithm, namely, compared with a compression algorithm which is not subjected to resolution enhancement processing, the combined processing algorithm provided by the invention enables the resolution of the recovered image to be remarkably enhanced.

There are many specific applications for remotely sensed images, which also require computer processing. In order to test the effect of the algorithm in the aspect of enhancing the resolution, the edge detection is carried out on the restored images obtained by the three compression algorithms, and the adopted edge detection operator is a general Prewitt operator.

Due to space limitation, the invention only gives the detection effect of the second local area. And edge detection is carried out on the original image of the second local area by using a Prewitt operator, and the number of detected edge points is 1437 points in total. Table 3-2 shows the edge detection results of the image restored by the Prewitt operator under different compression ratios for the three algorithms, where the detected edge points are true edge points (i.e. the detected edge points are always included in 1437 edge points of the original image) after the false edge points are removed, and the detection rate is the percentage of the number of edge points detected by the restored image to the number of edge points of the original image. As can be seen from the data in the table, under each compression ratio, the number of edge points and the detection rate of the restored image of the algorithm are better than those of the Jasper and Part 1 algorithms, and with the increase of the compression multiple, the detection rate of the algorithm is always kept above 90%, and the detection rate of the Jasper and Part 1 algorithms is reduced to below 90% when the compression multiple is nearly 27 times. In other words, for later edge detection application, the restored image when the algorithm is compressed to 64 times can also obtain better application effect than that when the Jasper and Part 1 algorithms are compressed to 27 times.

TABLE 3-2 Prewitt detection of edge points and detection Rate of restored images

Fig. 10, 11 to 13, and 14 to 16 are diagrams showing the effect of Prewitt edge detection on the second local area original and on the restored image after 16-fold and 64-fold compression, respectively. In the two comparative graphs, the detection results of the edge information such as the airplane and the house in fig. 13 and fig. 16 are obviously better than the edge detection results in fig. 11, fig. 12, fig. 14 and fig. 15, the edge continuity is better, and the effect is more remarkable particularly under a large compression ratio.

Therefore, the algorithm of the invention has better protection on the target edge information than the Jasper algorithm and the Part 1 algorithm, effectively enhances the resolution of the target edge, and enables the recovered image to be more beneficial to computer processing aiming at the later application.

On the premise of keeping the JPEG2000 standard compression flow basically unchanged, the algorithm of the invention is embedded with a multi-resolution information detection algorithm and a multi-resolution information protection algorithm, and the complexity of the encoder is increased to a certain extent, but the multi-resolution edge detection and the multi-resolution information protection algorithm adopted by the invention both utilize the original JPEG2000 technology as much as possible, and pay attention to reducing the calculated amount when designing a joint processing scheme, so the algorithm of the invention does not significantly increase the load of the encoder. This can be seen from the compression times of the three algorithms, which are shown in tables 3-3, for the algorithm of the present invention, which is between Jasper and Part 1. This is because the quantization step size of each sub-band set by the algorithm of the present invention is between Jasper and Part 1, so that the amount of data to be encoded by EBCOT is also between Jasper and Part 1, and the added extra calculation amount is not significant, so that the compression time is shorter than that of Jasper algorithm.

TABLE 3-3 comparison of compression time results (unit: ms)

Compression ratio	8	16	27	32	64
						Jasper	1078	1062	1053	1047	1040
Part 1	765	735	719	703	688
						Algorithm of the invention	906	891	884	879	875

Furthermore, another great advantage of the inventive algorithm is that it does not increase any burden on the decoder. The reason is that the joint processing algorithm does not change the main steps of the original JPEG2000 algorithm, only embeds the multiresolution information detection after the wavelet transformation, and adjusts the quantization step size of each sub-band and the weighting of the channel, because the quantization step size can be written into the header file of the code stream in an explicit mode, and the weighting value of the channel does not need to be used at the decoding end. Therefore, for the compressed code stream generated by the algorithm coding of the invention, any decoder supporting the JPEG2000 compressed code stream can be normally decompressed without changing the decoder.

Drawings

FIG. 1 is a remote sensing image to be processed in the present invention; FIGS. 2 and 3 are two local regions of the remote sensing image of FIG. 1, FIG. 2 being a first local region and FIG. 3 being a second local region; fig. 4 to 6 are restored images of the second local area by three algorithms at the time of 16-fold compression, fig. 4 is a restored image obtained by Jasper algorithm, fig. 5 is a restored image obtained by Part 1 algorithm, and fig. 6 is a restored image obtained by the algorithm of the present invention; fig. 7 to 9 are restored images of the second local area by three algorithms at the time of 32-fold compression, fig. 7 is a restored image obtained by Jasper algorithm, fig. 8 is a restored image obtained by Part 1 algorithm, and fig. 9 is a restored image obtained by the algorithm of the present invention; FIG. 10 is the Prewitt detection result of the second local area; fig. 11 to 13 are Prewitt edge detection effects of the restored image compressed by 16 times, where fig. 11 is that the Jasper algorithm obtains an edge detection effect, fig. 12 is that the Part 1 algorithm obtains an edge detection effect, and fig. 13 is that the algorithm of the present invention obtains an edge detection effect; fig. 14 to 16 are Prewitt edge detection effects of the 64 times compressed restored image; wherein, fig. 14 shows that the Jasper algorithm obtains the edge detection effect, fig. 15 shows that the Part 1 algorithm obtains the edge detection effect, and fig. 16 shows that the algorithm obtains the edge detection effect; FIG. 17 is a flow chart of the present invention; FIGS. 18 and 19 are type weighting values of different sub-bands of a wavelet transform, where FIG. 18 is a one-layer decomposition and FIG. 19 is a two-layer decomposition; fig. 20 is a correspondence relationship of wavelet coefficients in adjacent resolution subbands.

Detailed Description

In the first embodiment, the present embodiment is described with reference to fig. 17, and the steps of the present embodiment are as follows:

the method comprises the following steps: inputting an image f (x, y) to be compressed;

step two: preprocessing the image;

the preprocessing comprises image slicing, DC level translation and component transformation;

image slices operate on large images, and are formed by dividing an image into rectangular blocks, namely image slices (tiles), which do not overlap with each other, and performing compression coding on each image slice as an independent image.

DC level shifting is only performed for images consisting of unsigned numbers. The same level is subtracted for each sample. If the pixel value is p bits, subtract 2^p-1The original range is made to be [0, 2 ]^p-1]Is shifted to-2^p-1，2^p-1-1]Within the range. For example, 8bit level values 0-255, 128 is subtracted per pixel value. The component transformation is operated on color images in order to reduce the correlation between the components.

There is a strong correlation between the color components, e.g., R, G, B, of a multi-component image. The R, G, B three components are converted into luminance chrominance components Y, Cr, Cb by a component transform.

Step three: performing discrete wavelet transform on the preprocessed data;

discrete wavelet transform uses Mallat decomposition to perform a series of high-pass and low-pass filtering on the source signal. After each filtering, the data sampling frequency is reduced to half of the original frequency, so as to ensure that the number of the transformed coefficients is the same as that of the source signals. The low-frequency information of the source information is output by low-pass filtering each time, the source signal is reproduced by lower resolution, and most energy of the source signal is concentrated; and the high-pass filtering outputs high-frequency information with less energy. After one low-pass filtering, the signal still has strong correlation and needs to be filtered again. While the high frequency signal correlation is already weak, it is not cost effective to filter it again.

The discrete wavelet transform is implemented using an Le Gall 5/3 filter or a Daubechies 9/7 filter. The Le Gall 5/3 filter is integer type for lossy or lossless compression; the latter Daubechies 9/7 filter is floating point type and can only be used for lossy compression. These two filter coefficients are given in tables 2-1 and 2-2.

TABLE 2-1 Le Gall 5/3 analysis and Synthesis of Filter coefficients

Table 2-2Daubechies 9/7 analysis and synthesis of filter coefficients

Step four: carrying out information detection on the result of the wavelet transform;

the adopted information detection method is a multi-resolution edge detection algorithm and is divided into the following 4 steps, wherein the step 1 and the step 2 are based on the band-pass characteristic of wavelet decomposition, and three high-frequency sub-bands LH of the wavelet decomposition are extracted for the purpose of extracting the image edge due to the fact that the edge exists in the high-frequency component_j、HL_jAnd HH_jTaking a threshold to generate respective edge images E _ LH_j、E_HL_j、E_HH_j；

Step 1: self-adaptively setting double thresholds TH of each sub-band_b1And TH_b2；

According to the self characteristics of each high-frequency sub-band under different resolutions, self-adaptively setting double thresholds, first threshold TH_b1Being the absolute median of the sub-band b, sub-threshold TH_b2First threshold value TH_b1One half of (a);

${\mathrm{<figure-callout\; id="1"\; label="TH\; b"\; filenames="HSA00000108778600034.png"\; state="\{\{state\}\}">TH</figure-callout>}}_b=\frac{1}{2}\max \left(\right|w_b(i,j)\left|\right)$

${\mathrm{TH}}_{b2}=\frac{1}{2}{\mathrm{<figure-callout\; id="1"\; label="TH\; b"\; filenames="HSA00000108778600034.png"\; state="\{\{state\}\}">TH</figure-callout>}}_b=\frac{1}{4}\max \left(\right|w_b(i,j)\left|\right)$

wherein, w_b(i, j) are the wavelet coefficient values of subband b, the sign "| w_b(i, j) | "represents taking the absolute value of w (i, j);

step 2: detecting multi-resolution edge information by a double-threshold method;

the wavelet coefficients of the high-frequency sub-bands are judged as follows:

if w_b(i，j)＞TH_b1If yes, then (i, j) is judged as the edge point of the sub-band b;

if w_b(i，j)＜TH_b1But w_b(i，j)＞TH_b2Then search the wavelet coefficient in 8 fields to see if there is more than the first threshold TH_b1The presence of a coefficient of (a); if the edge point exists in 8 fields, the probability that the edge point is also the edge point is very high, so that (i, j) is determined to be the edge point; if not, it indicates that no edge point exists in the 8 fields, and the probability that the point is an edge point is very small, so that (i, j) is determined to be a non-edge point;

if w_b(i，j)＜TH_b2If yes, then (i, j) is judged as a non-edge point;

thus, for each resolution, three edge images E _ LH containing different directivities are obtained_j、E_HL_j、E_HH_j；

And step 3: matching and denoising edge information with different resolutions;

edge image E _ LH of different resolution generated by high frequency component using wavelet decomposition in previous step_j、E_HL_j、E_HH_jIt is sensitive to noise and needs to be processed to suppress noise.

Because the true edge points of the image have larger energy in the wavelet representation of each layer of decomposition and have stronger correlation; the energy distribution of the noise in each channel is different and is not related to each other. Therefore, with the multi-resolution decomposition characteristic of wavelet transform, the noise suppression method of cross energy intersection is adopted: for the generated edge image E, the original edge signal property (positive or negative) is kept unchanged_jAnd performing cross processing on two (or more) adjacent layers to generate a new edge image Ee according to the following formula:

E_e＝sgn{E_j}×|E_j|×|E_j+1|

although the cross energy cross processing method of two adjacent layers can effectively remove noise, the calculation amount is too large because the two adjacent layers need to be multiplied. The principle based on which the cross energy intersection method is based is strong correlation of edge information and non-correlation of noise, and based on the same principle, the same noise suppression effect can be achieved by matching the edge information in each sub-band under different resolutions, and meanwhile, the calculation amount can be greatly reduced.

And C, according to the wavelet transformation principle, carrying out position matching on the edge points of the second high resolution and the highest resolution detected in the step II. Find out 4 points (2i, 2j), (2i, 2j +1), (2i +1, 2j) and (2i +1, 2j +1) corresponding to the edge point (i, j) of the second highest resolution under the adjacent highest resolution, as shown in fig. 20, determine whether there is at least one edge point in the 4 points. If yes, judging that all the points on the matching are real edge points in the multi-resolution edge information matching; if not, the matching is ended, the point is indicated as noise, and the qualification of the edge point is cancelled. And when all the edge points with the second highest resolution are matched, judging the edge points which are not matched with the highest resolution as noise and discarding the noise.

And 4, step 4: extracting the edge information of the interested target;

the edge information detected through the above steps includes not only the target edge of interest, but also background edges that are not of interest to the user, such as edges at sea-sky boundaries, texture edges on the ground, and the like. If all the detected edge information is stored, on the premise that the target code rate is fixed, corresponding other information must be discarded, which may cause a large reduction in the quality of the restored image, especially when the information discarded by replacement is low-frequency information. In addition, it is meaningless to reserve background edges which are not interesting to the user and are not concerned by the later application, so that it is necessary to select the detected edges and extract target edges which are interesting to the user or are relatively concerned by the later application.

Some important information in the remote sensing image is mainly hidden in points, lines and point sets, such as tanks, airplanes, bridges, houses and the like, and usually only accounts for dozens of pixels to hundreds of pixels. And different from the background edge, the edge information of the target is concentrated, has certain geometric shape characteristics, and has a large range of closed or approximately closed shape for the edge of a larger target. Thus, the object of interest may be determined according to the intensity of the edge information, the geometric features, or the closeness of the edges. In order to reduce the increased amount of computation as much as possible, the edges of the object of interest are extracted according to the intensity of the edge information.

Extracting edges of the object of interest according to the intensity of the edge information;

setting different dimensions of window W for different resolutions_jThe difference between the window length and width 1/2 of adjacent resolutions, i.e. W_j+1＝W_jSelecting the window with the highest resolution as the code block, so that the window size is matched with the resolution of the window;

searching edge information dense areas corresponding to different resolutions by using respective windows of each scale;

the three high-frequency subbands of the highest resolution are first searched in detail, due to the high-frequency subband LH_j、HL_jAnd HH_jRespectively contain edge information in the horizontal direction, the vertical direction and the diagonal direction, and thus, the windowThe port scan order is: LH_jSubband horizontal scanning, HL_jSub-band vertical scanning, HH_jSub-band horizontal or vertical scanning;

then, the searching result of the window with the highest resolution guides the searching of three high-frequency sub-bands with the second highest resolution;

counting edge points NE within a window_biFinding out the most dense area of each sub-band edge and recording the number of edge points MaxNE_bWhen NE_bi＞MaxNE_bWhen the code block is positioned at the position of the target, 2, the edge of the target of interest is marked, and the code block where the target of interest is positioned is also marked; the remaining edge points are considered as background edges. Thus, most of the background edges can be effectively filtered out. While still some background edges exist, for the application of the present invention, a certain false alarm probability is allowed.

Step five: carrying out quantization processing on the wavelet transform result and simultaneously carrying out sub-band weighting;

the quantization and weighting method is as follows:

quantization is a process that reduces the accuracy of transform coefficients. If the quantization step is not 1, the quantization is lossy. With uniform quantization, all wavelet transform coefficients a in sub-band b_b(u, v) are both quantized to q according to_b(u，v)：

Wherein

R_bIs the nominal dynamic range, ε, of sub-band b_bAnd mu_bRespectively representing the exponent and the false number of the corresponding sub-band b; r_b、ε_b、μ_bDifferent values may be available for different sub-bands;

in order to save multi-resolution edge information during quantization, on the basis of a quantization step setting method of a JPEG2000Jasper algorithm, the quantization step of each sub-band is integrally adjusted: HH with highest resolution₁The sub-band is used as a reference, the quantization step size is set to be 1, and the relative proportion relation between the rest sub-bands and the sub-band is kept unchanged for adjustment. After adjustment, the quantization step size of all sub-bands is less than or equal to 1, so that the detected multi-resolution edge information is ensured not to be quantized and lost. The same subband has only one quantization step, but each subband may have a different quantization step, as shown in table 1-1:

TABLE 1-1 adjusted subband quantization step size

Step six: and E, entropy coding is carried out on the quantization result obtained in the step five, and a layer one coding method is adopted in an EBCOT coding method.

Step seven: carrying out code stream truncation on the entropy coding result, and simultaneously carrying out channel weighting;

carrying out three-time channel scanning and coding on each bit plane from the MSB (most significant bit) to the LSB (least significant bit) of the quantized wavelet coefficients of each sub-band to obtain an embedded code stream;

finding out channels where all corresponding multi-resolution edge information is located according to the code blocks where the edges of the interested target marked in the step four are located, and weighting the channels;

the influence of different resolution components on system errors, the visual characteristics of human eyes, the characteristics of the channels and other factors are comprehensively considered, and the weighting formula for setting the channels is as follows:

the weighting formula for the setup channel is as follows:

<math><mrow><mi>&Delta;</mi><msubsup><mi>D</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup><mo>=</mo><msub><mi>EP</mi><mi>b</mi></msub><mo>&times;</mo><msub><mi>T</mi><mi>b</mi></msub><mo>&times;</mo><mroot><msub><mi>&gamma;</mi><mi>b</mi></msub><mn>4</mn></mroot><mo>&times;</mo><msup><mn>2</mn><mrow><mn>2</mn><mi>p</mi></mrow></msup><mo>&times;</mo><msubsup><mi>M</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup></mrow></math>

<math><mrow><msub><mi>EP</mi><mi>b</mi></msub><mo>=</mo><mfenced open='{' close=''><mtable><mtr><mtd><mfrac><mrow><msub><mi>T</mi><mi>LL</mi></msub><mo>&times;</mo><mroot><msub><mi>&gamma;</mi><mi>LL</mi></msub><mn>4</mn></mroot></mrow><mrow><msub><mi>T</mi><mi>b</mi></msub><mo>&times;</mo><mroot><msub><mi>&gamma;</mi><mi>b</mi></msub><mn>4</mn></mroot></mrow></mfrac></mtd><mtd><msub><mi>P</mi><mi>bi</mi></msub><mo>&Element;</mo><mi>EP</mi></mtd></mtr><mtr><mtd><mn>1</mn></mtd><mtd><msub><mi>P</mi><mi>bi</mi></msub><mo>&NotElement;</mo><mi>EP</mi></mtd></mtr></mtable></mfenced></mrow></math>

wherein,

is the distortion factor after the channel weighting,

is code block B_iN of (2)_iThe channel self information quantity corresponding to the cut-off point, p is the bit plane of the channel, gamma_bIs the energy weighting of the sub-band b, T_bThe weighted value is the type weighted value of different sub-bands, is irrelevant to the image decomposition layer number and is only relevant to the sub-band type; as shown in fig. 18 and 19; EP_bFor the weighting values of the edge channels, EP denotes the set of all multi-resolution edge channels, P_biI channel of sub-band b, T_LLAnd gamma_LLRespectively representing the type weighted value and the energy weighted value of the LL sub-band; thus, the distortion factor after the channel weighting can be obtained

Then, the rate distortion slope value of each channel is obtained according to the following formula

<math><mrow><msubsup><mi>S</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup><mo>=</mo><mfrac><msubsup><mi>&Delta;D</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup><msubsup><mi>&Delta;L</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup></mfrac><mo>=</mo><mfrac><mrow><msubsup><mi>D</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup><mo>-</mo><msubsup><mi>D</mi><mi>i</mi><mrow><msub><mi>n</mi><mi>i</mi></msub><mo>-</mo><mn>1</mn></mrow></msubsup></mrow><mrow><msubsup><mi>L</mi><mi>i</mi><mrow><msub><mi>n</mi><mi>i</mi></msub><mo>-</mo><mn>1</mn></mrow></msubsup><mo>-</mo><msubsup><mi>L</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup></mrow></mfrac></mrow></math>

With the above weighting, the multiresolution edge channel can achieve approximately the same importance as the LL subband, i.e., the multiresolution edge channel is nearly as important as the channel in the same bit plane in the LL subband and more important than the other non-edge channels in the same subband. Therefore, according to the rate-distortion optimization criterion, when in truncation, the multi-resolution edge channel and the channel in the LL subband are preferentially reserved, so that the combined processing of high-fidelity remote sensing image compression and resolution enhancement is realized.

Step eight: and obtaining a compressed code stream.

Claims (7)

1. The combined processing method for high-fidelity remote sensing image compression and resolution enhancement is characterized by comprising the following steps of:

the method comprises the following steps: inputting an image f (x, y) to be compressed;

step two: preprocessing the image;

step three: performing discrete wavelet transform on the processed data;

step four: carrying out information detection on the result of the wavelet transform;

step five: carrying out quantization processing on the wavelet transform result and simultaneously carrying out sub-band weighting;

step six: entropy coding is carried out on the quantization result of the step five;

step seven: carrying out code stream truncation on the entropy coding result, and simultaneously carrying out channel weighting;

step eight: and obtaining a compressed code stream.

2. The joint processing method for high-fidelity remote sensing image compression and resolution enhancement as claimed in claim 1, wherein the image preprocessing in the second step comprises image slicing, DC level translation and component transformation.

3. The joint processing method for high-fidelity remote sensing image compression and resolution enhancement as claimed in claim 1, characterized in that the discrete wavelet transform in step three is implemented by using Le Gall 5/3 filter or Daubechies 9/7 filter.

4. The high-fidelity remote sensing image compression and resolution enhancement combined processing method according to claim 1, characterized in that the information detection method adopted in the fourth step is a multi-resolution edge detection algorithm, and the multi-resolution edge detection algorithm is divided into the following steps:

step 1: self-adaptively setting double thresholds TH of each sub-band_b1And TH_b2；

According to the self characteristics of each high-frequency sub-band under different resolutions, self-adaptively setting double thresholds, first threshold TH_b1Being the absolute median of the sub-band b, sub-threshold TH_b2First threshold value TH_b1One half of (a);

${\mathrm{TH}}_{b1}=\frac{1}{2}\max \left(\right|w_b(i,j)\left|\right)$

${\mathrm{TH}}_{b2}=\frac{1}{2}{\mathrm{TH}}_{b1}=\frac{1}{4}\max \left(\right|w_b(i,j)\left|\right)$

wherein, w_b(i, j) are the wavelet coefficient values of subband b, the sign "| w_b(i, j) | "represents taking the absolute value of w (i, j);

step 2: detecting multi-resolution edge information by a double-threshold method;

the wavelet coefficients of the high-frequency sub-bands are judged as follows:

if w_b(i，j)＞TH_b1If yes, then (i, j) is judged as the edge point of the sub-band b;

if w_b(i，j)＜TH_b1But w_b(i，j)＞TH_b2Then search the wavelet coefficient in 8 fields to see if there is more than the first threshold TH_b1The presence of a coefficient of (a); if yes, judging (i, j) as an edge point; if not, (i, j) is judged as a non-edge point;

if w_b(i，j)＜TH_b2If yes, then (i, j) is judged as a non-edge point;

thus, for each resolution, three edge images E _ LH containing different directivities are obtained_j、E_HL_j、E_HH_j；

And step 3: matching and denoising edge information with different resolutions;

and 4, step 4: extracting the edge information of the interested target;

extracting edges of the object of interest according to the intensity of the edge information;

setting different dimensions of window W for different resolutions_jThe difference between the window length and width 1/2 of adjacent resolutions, i.e. W_j+1＝W_jSelecting the window with the highest resolution as the code block;

using each kind of rulerSearching an edge information dense area corresponding to the resolution ratio in the degree window; the search is performed first for the three high-frequency subbands of the highest resolution, due to the high-frequency subband LH_j、HL_jAnd HH_jContains edge information in the horizontal, vertical and diagonal directions, respectively, and therefore the scanning order of the windows is: LH_jSubband horizontal scanning, HL_jSub-band vertical scanning, HH_jSub-band horizontal or vertical scanning; then, the searching result of the window with the highest resolution guides the searching of three high-frequency sub-bands with the second highest resolution;

counting edge points NE within a window_biFinding out the most dense area of each sub-band edge and recording the number of edge points MaxNE_bWhen NE_bi＞MaxNE_bWhen the code block is positioned at the position of the target, 2, the edge of the target of interest is marked, and the code block where the target of interest is positioned is also marked; the remaining edge points are considered as background edges.

5. The joint processing method for compressing and enhancing resolution of the high-fidelity remote sensing image according to claim 1, wherein the quantization step size of each sub-band in the fifth step is as shown in table 1-1:

6. the joint processing method for high-fidelity remote sensing image compression and resolution enhancement as claimed in claim 1, characterized in that the entropy coding in the sixth step adopts an EBCOT layer-coding method.

7. The high-fidelity remote sensing image compression and resolution enhancement combined processing method according to claim 4, characterized in that the seventh step of performing code stream truncation on the entropy coding result and simultaneously performing channel weighting comprises the following steps:

carrying out three-time channel scanning and coding on each bit plane from the most significant bit to the least significant bit of the wavelet coefficients of each quantized sub-band to obtain an embedded code stream;

finding out channels where all corresponding multi-resolution edge information is located according to the code blocks where the edges of the interested target marked in the step 4 are located in the step four, and weighting the channels;

the weighting formula for the setup channel is as follows:

<math><mrow><mi>&Delta;</mi><msubsup><mi>D</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup><mo>=</mo><msub><mi>EP</mi><mi>b</mi></msub><mo>&times;</mo><msub><mi>T</mi><mi>b</mi></msub><mo>&times;</mo><mroot><msub><mi>&gamma;</mi><mi>b</mi></msub><mn>4</mn></mroot><mo>&times;</mo><msup><mn>2</mn><mrow><mn>2</mn><mi>p</mi></mrow></msup><mo>&times;</mo><msubsup><mi>M</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup></mrow></math>

<math><mrow><msub><mi>EP</mi><mi>b</mi></msub><mo>=</mo><mfenced open='{' close=''><mtable><mtr><mtd><mfrac><mrow><msub><mi>T</mi><mi>LL</mi></msub><mo>&times;</mo><mroot><msub><mi>&gamma;</mi><mi>LL</mi></msub><mn>4</mn></mroot></mrow><mrow><msub><mi>T</mi><mi>b</mi></msub><mo>&times;</mo><mroot><msub><mi>&gamma;</mi><mi>b</mi></msub><mn>4</mn></mroot></mrow></mfrac></mtd><mtd><msub><mi>P</mi><mi>bi</mi></msub><mo>&Element;</mo><mi>EP</mi></mtd></mtr><mtr><mtd><mn>1</mn></mtd><mtd><msub><mi>P</mi><mi>bi</mi></msub><mo>&NotElement;</mo><mi>EP</mi></mtd></mtr></mtable></mfenced></mrow></math>

wherein,

is the distortion factor after the channel weighting,is code block B_iN of (2)_iThe channel self information quantity corresponding to the cut-off point, p is the bit plane of the channel, gamma_bIs the energy weighting of the sub-band b, T_bWeighting values for different sub-bands; EP_bFor the weighting values of the edge channels, EP denotes the set of all multi-resolution edge channels, P_biI channel of sub-band b, T_LLAnd gamma_LLRespectively representing the type weighted value and the energy weighted value of the LL sub-band; thereby obtaining the distortion degree of the channel after weighting

Then, the rate distortion slope value of each channel is obtained according to the following formula

<math><mrow><msubsup><mi>S</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup><mo>=</mo><mfrac><msubsup><mi>&Delta;D</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup><mrow><mi>&Delta;</mi><msubsup><mi>L</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup></mrow></mfrac><mo>=</mo><mfrac><mrow><msubsup><mi>D</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup><mo>-</mo><msubsup><mi>D</mi><mi>i</mi><mrow><msub><mi>n</mi><mi>i</mi></msub><mo>-</mo><mn>1</mn></mrow></msubsup></mrow><mrow><msubsup><mi>L</mi><mi>i</mi><mrow><msub><mi>n</mi><mi>i</mi></msub><mo>-</mo><mn>1</mn></mrow></msubsup><mo>-</mo><msubsup><mi>L</mi><mi>i</mi><msub><mi>n</mi><mi>i</mi></msub></msubsup></mrow></mfrac></mrow></math> 。

Images