CN111027511B - Remote sensing image ship detection method based on region of interest block extraction - Google Patents

Abstract

The invention discloses an optical remote sensing image ship detection method based on the extraction of interesting blocks, which mainly solves the problems of low detection accuracy and many false alarms in the prior art. Its implementation plan is: constructing optical remote sensing image ship detection data set; downsampling and dehazing enhancement of wide remote sensing image, using context information and image global features to perform water and land segmentation; using the constructed data set to train a target detection model based on SCRDet ; According to the result of water and land segmentation, scan the original wide-format remote sensing image with a partially overlapping sliding window to extract the block of interest as the area to be detected, input the image of the to-be-detected area into the detection model to obtain the regional detection result; map the regional result to the original wide-format image On the scale, improved non-maximum suppression is used to optimize the initial detection results; the detection results are optimized again according to the structural characteristics of the ship. The invention has high detection accuracy and low false alarm rate, and can be used for acquiring interesting ship targets and their positions in large-scale remote sensing images.

Description

Remote sensing image ship detection method based on extraction of interesting blocks

technical field

The invention belongs to the technical field of image processing, and in particular relates to a method for detecting ships in optical remote sensing images, which can be used for target recognition in large-format remote sensing images.

Background technique

Target detection of optical remote sensing images is one of the important issues in the field of remote sensing image research, and ship target detection has extremely important application value in fishery management, military reconnaissance and strategic deployment due to its particularity and criticality. Ship target detection is to determine whether there is a ship in the water or on the shore from a complex scene, and to locate it.

Traditional ship target detection methods mainly include methods using land and sea segmentation and using prior geographic information. These methods need to manually design a large number of features to locate ships. The adaptability is not strong when the target is detected at the same time.

In recent years, deep learning has developed rapidly, and deep convolutional neural networks have achieved advanced results in the field of image processing because of their ability to automatically extract shallow and deep features of images, avoiding the manual feature extraction of traditional methods. Currently, methods based on deep learning are the mainstream methods for object detection.

The ship target detection method based on deep learning can be divided into two categories according to the detection frame: based on the horizontal detection frame and based on the tilt detection frame. in:

Methods based on horizontal detection boxes include the classic two-stage detection framework Faster-RCNN, and the single-stage detection frameworks Yolo, RetinaNet, and SSD.

Methods based on tilt detection boxes, including R ² CNN, ROI-Transformer, SCRDet, etc.

Due to the characteristics of arbitrary ship direction and dense ships docking, the tilt detection frame can more accurately locate the ship. In the method of using the inclined detection frame, the SCRDet (Towards More Robust Detection for Small, Cluttered and Rotated Objects) model can fully integrate the low-level features and high-level features, and achieve better detection results for dense objects. However, deep learning-based methods do not perform water and land segmentation on images, but directly input wide image slices into the trained model to generate results. Inputting complex land areas without water into the detection model not only reduces the detection efficiency, but also reduces the detection efficiency. It may cause obvious false alarms on land. In addition, the incomplete hull caused by the segmented image will also cause obvious missed detection, so the detection accuracy is not high.

SUMMARY OF THE INVENTION

The purpose of the present invention is to propose a ship detection method based on an optical remote sensing image extraction based on the block of interest, in order to optimize the positioning of the ship target in the remote sensing image and improve the detection efficiency and accuracy, aiming at the above-mentioned shortcomings of the prior art.

To achieve the above object, the technical scheme of the present invention comprises the following steps:

(1) Constructing the optical remote sensing image ship detection dataset G:

1a) Download the optical remote sensing data of Gaofen-2, manually filter out the areas that contain ship targets, crop and save these areas in a partially overlapping manner;

1b) Randomly flip or rotate all the images obtained in 1a) up, down, left and right to obtain an amplified image and save it;

1c) Annotate the amplified image with a slanted rectangular frame, save the annotation information as an xml format file, and use all the amplified images and their corresponding annotation information to form the optical remote sensing image ship detection dataset G;

(2) down-sampling the wide-format optical remote sensing image, and using the dehazing algorithm based on the dark channel prior to reduce the cloud and fog occlusion on the image, and obtain the enhanced image I;

(3) Use the context information and image global features to perform water and land segmentation on the enhanced image I, extract the water area of interest, and obtain a binary image:

(3a) using the context feature information of the pixel to perform preliminary threshold segmentation on the enhanced image I to obtain a preliminary water and land segmentation binary map I ₃ ;

(3b) Mark the water-connected domain set W _{= { w 1} , _. The set _{L = { l 1 ,} . The regional characteristics of the domain and the land-connected domain are re-calibrated according to the discrimination rules, and the optimized water-land segmentation binary map I ₄ of the enhanced image I is obtained;

(3c) after performing the morphological expansion operation on the optimized water and land segmentation binary image I ₄ , the upsampling is restored to the original wide image size, and the final water and land segmentation binary image I ₅ is obtained;

(4) Using random multi-scale strategy to train the SCRDet target detection model M ₀ based on convolutional neural network, and obtain the trained detection model M;

(5) Using the final water and land segmentation binary image I ₅ , use a partially overlapping sliding window to extract the block of interest as the area to be detected on the original wide image, and form the image set of the area to be detected as F={f ₁ ,... , f _i ,...,f _n }, the position set is S={s ₁ ,...,s _i ,...,s _n }, where f _i represents the ith detection area, and s _i represents the detection area The coordinates of the upper left corner of the area _fi , input the image set F of the area to be detected into the detection model M, and obtain the area detection result of each area _fi ;

其中Pⁱ、

分别表示初步检测的第i个检测框的类别、位置坐标、置信度分数；对初步检测结果集合A作改进的非极大值抑制，得到一次优化检测结果集合

其中Qⁱ、

分别表示一次优化的第i个检测框的类别、位置坐标、置信度分数；(6) Map the region detection results to the original wide image scale according to the position set S, and obtain a preliminary detection result set

where P ⁱ ,

Respectively represent the category, position coordinates, and confidence score of the ith detection frame of the preliminary detection; perform improved non-maximum suppression on the preliminary detection result set A, and obtain an optimized detection result set

where Q ⁱ ,

Respectively represent the category, position coordinates, and confidence score of the ith detection frame optimized once;

其中Rⁱ、

分别表示最终生成的的第i个检测框的类别、位置坐标、置信度分数。(7) Perform secondary optimization on the primary optimization detection result set B according to the structural characteristics of the ship, and obtain the final detection result set

where R ⁱ ,

Represent the category, position coordinates, and confidence score of the finally generated i-th detection frame, respectively.

Compared with the prior art, the present invention has the following advantages:

1. The present invention uses the context information and global features of the enhanced image that is downsampled and dehazed to perform water and land segmentation according to the complex and large size of the wide-format remote sensing image scene, and performs morphological operations, which can more effectively optimize the water and land segmentation results. , retains ship targets within the block of interest.

2. The present invention scans the whole image through partially overlapping sliding windows, and extracts the block of interest as the area to be detected according to the result of the water and land segmentation, which effectively improves the target detection efficiency while reducing the false alarm on the land; The improved non-maximum value suppression of the detection results on the wide scale can effectively reduce the missed detection caused by the image slice and improve the detection accuracy.

Description of drawings

Fig. 1 is the realization flow schematic diagram of the present invention;

Fig. 2 is the simulation result image of carrying out water and land segmentation to wide remote sensing image with the present invention;

Fig. 3 is the simulation result image of carrying out ship detection to wide-format remote sensing image by the existing method;

FIG. 4 is a simulation result image of ship detection on a wide-format remote sensing image using the present invention.

Detailed ways

The embodiments and effects of the present invention will be further described below with reference to the accompanying drawings.

1, the implementation steps of this embodiment are as follows:

Step 1, construct the optical remote sensing image ship detection dataset G.

1.1) Download the optical remote sensing data of Gaofen-2, manually filter out the areas containing the ship target, crop these areas with a partially overlapping sliding window with a size of 832×832 and a step size of 416 and save them;

1.2) Randomly flip or rotate all the images obtained in 1.1) up, down, left, and right to obtain an amplified image and save it;

1.3) Mark all augmented images obtained in 1.2) with a slanted rectangular frame, save the annotation information as an xml format file, and use all augmented images and their corresponding annotation information to form an optical remote sensing image ship detection dataset G .

Step 2, the wide optical remote sensing image is down-sampled and enhanced by dehazing.

The original wide-format optical remote sensing image is down-sampled by 8 times. According to the dehazing algorithm based on dark channel prior proposed by He Yuming in "Single Image HazeRemoval Using Dark Channel Prior", the cloud and fog occlusion on the original wide-format optical remote sensing image is reduced. , the enhanced image I is obtained.

Step 3, using context information and global image features to perform water and land segmentation.

3.1) utilize the context feature information of the pixel to make preliminary threshold segmentation to the enhanced image I;

3.1.a) Extend the edge of the enhanced image I with a mirror image with a width of 3 to obtain the extended image I', take each pixel of the enhanced image I as the center of a 7 × 7 sliding window, and extract the corresponding window on the extended image I' region and calculate the regional variance v as the context feature value of the central pixel point to obtain the feature map V of the enhanced image I, and the feature value of the enhanced image I at the coordinates (i, j) is V(i, j);

3.1.b) Count the distribution histogram of the feature map V with 1 as the interval, determine the peak variance v _max corresponding to the peak of the histogram, and calculate the value corresponding to the first valley value of the histogram in the range of [v _max ,+∞) valley variance g _t ;

3.1.c) Compare the eigenvalues V(i,j) of the enhanced image I at coordinates (i,j) with the valley variance g _t :

If V(i,j)≤g _t , then update V(i,j)=1,

Otherwise, update V(i,j)=0;

3.1.d) perform a morphological expansion operation on the updated feature map V to obtain a feature binary map I ₁ ;

对增强图像I作拉普拉斯灰度变换，利用OTSU自适应阈值法分割灰度变换后的增强图像I，得到灰度二值图I₂；3.1.e) Using the modified Laplacian

Laplacian grayscale transformation is performed on the enhanced image I, and the enhanced image I after the grayscale transformation is segmented by using the OTSU adaptive threshold method to obtain a grayscale binary image I ₂ ;

3.1.f) The feature binary image I ₁ and the gray-scale binary image I ₂ are correspondingly performed pixel-by-pixel logical AND operation to obtain a preliminary water and land segmentation binary image I ₃ , wherein the pixel value 0 in I ₃ represents land, and the pixel value 1 means water;

3.2) Using the characteristics of the water connectivity domain and the land connectivity domain, the category re-calibration of the connectivity domain is carried out:

_3.2.a ) Mark the water connected domain set W= _{{ w 1} , _. Calculate the area of the n connected water areas, take the three connected areas with the largest area as the water feature template of the wide image, and calculate the variance σ ² of the water feature template;

3.2.b) Initialize the area threshold t ₁ =35×35;

和方差

若

且

则重标定w_i为水域，否则，重标定w_i为陆地；3.2.c) Traverse the water connectivity domain set W, and calculate the area of w _i in W

and variance

like

and

Then re-calibrate _wi as water, otherwise, re-calibrate _wi as land;

3.2.d) Mark the land-connected domain set L={l ₁ ,...,l _j ,...,l _m } in the preliminary water-land segmentation binary map _I3 , where l _j represents the j-th land-connected domain;

若

则重标定l_j为水域，否则，重标定l_j陆地；3.2.e) Traverse the land-connected domain set L, and calculate the area of l _j in L as

like

Then re-calibrate l _j as water, otherwise, re-calibrate l _j land;

3.2.f) assign values 0 and 1 to all pixels in the re-calibrated land-connected domain and water-connected domain, respectively, to obtain an optimized water-land segmentation binary map I ₄ ;

3.3) Using the optimized binary map I ₄ of water and land segmentation to obtain the final binary map I ₅ of water and land segmentation:

3.3.a) Copy the optimized binary image I ₄ of land and water segmentation to obtain the replicated result image I ₄ ′;

3.3.b) Use 15×15 ellipse structural elements to scan and optimize the binary image I ₄ of water and land segmentation pixel by pixel, and perform logical AND operation between the structural elements and each pixel in the corresponding area of I ₄ :

3.3.c) Judging the result of each pixel point, update the center pixel in the corresponding area of I ₄ ':

If the result of each pixel is 0, update the center pixel of the corresponding area to 0;

Otherwise, update the center pixel of the corresponding area to 1;

3.3.d) Upsampling the updated I ₄ ' by 8 times, restoring the size of I ₄ ' to the original wide image size, and obtaining the final binary image I ₅ of land and water segmentation.

Step 4, random multi-scale training of the SCRDet model M ₀ based on the convolutional neural network.

4.1) Take 90% of the optical remote sensing image ship detection dataset G constructed in step 1 as the training sample, and the remaining 10% as the test sample;

4.2) Use the ResNet-50 network as the backbone network of the SCRDet model M ₀ , and use the dataset ImageNet to pre-train the model parameters obtained by the ResNet-50 network, and then use the model parameters to initialize the parameters of the backbone network of the SCRDet model M ₀ ;

4.3) Randomly select a training image and its corresponding label in the training sample, and randomly flip the image and label up and down or left and right to obtain the transformed image X and the real label Y;

4.4) Randomly select a certain scale from [600, 700, 832] as the short side length of the image, scale the transformed image X and the ground truth Y to the selected scale, and input it into the SCRDet model M ₀ to get the prediction result y';

4.5) Calculate the error between the predicted result Y' and the real label Y, and use the optimizer Adam to minimize the error to update the weight parameter of the SCRDet model M ₀ ;

4.6) Repeat 4.3)-4.5), when the number of training rounds reaches 200,000, the trained target detection model M is obtained.

Step 5: Extract the area to be detected on the original wide image, and input it into the trained target detection model M to detect the area.

5.1) Initialize the window parameter p=832, and initialize the image set F and the position set S of the area to be detected to be an empty set;

5.2) Scan the original wide-scale remote sensing image with a partially overlapping sliding window with a size of p × p and a step size of p/2, and calculate the final water and land segmentation binary map I ₅ area corresponding to the window area f of the original wide-scale remote sensing image The water area S _w within;

5.3) According to the size of the water area S _w , determine whether the window area f is a block of interest:

If S _w >(p/8) ² , the window area f is determined to be the block of interest, the window area f is added to the image set F of the area to be detected, and the upper left corner coordinate s of the window area f is added to the position Set S, obtain the image set F={f ₁ ,...,f _i ,...,f _n } and the position set S={s ₁ ,...,s _i ,...,s _n };

Otherwise, it is determined that the window area f is not a block of interest, and no operation is performed;

5.5) Input the image set F={f ₁ ,...,f _i ,...,f _n } into the detection model M to obtain the category and position coordinates of all detection frames of each area f _i , the confidence score.

Step 6: Perform improved non-maximum suppression on the preliminary detection results on the scale of the original wide image.

6.1) According to the coordinates s _i of the area f _i to be detected in the position set S={s ₁ ,...,s _i ,...,s _n }, map the detection result of each area f _i to the original width At the image scale, the initial detection result set is obtained:

分别表示初步检测的第i个检测框的类别、位置坐标、置信度分数；where P ⁱ ,

Respectively represent the category, position coordinates, and confidence score of the i-th detection frame of the preliminary detection;

6.2) Initialize the optimized detection result set B to be empty;

根据索引r从A中取出检测结果

加入集合B后，并将

从集合A中删除；6.3) Assume that the categories of all detection frames in the preliminary detection result set A are the same category, and set the index of the detection result with the highest confidence in A

Take the detection result from A according to the index r

After joining set B, and put

remove from set A;

即两个检测框重叠区域面积与并集区域面积的比值；6.4) Calculate the intersection ratio of each detection frame in the preliminary detection result set A and the detection frame with the highest confidence

That is, the ratio of the area of the overlapping area of the two detection frames to the area of the union area;

与交并比阈值t₂的大小关系：6.5) Judging the intersection ratio

The size relationship with the intersection and union ratio threshold t ₂ :

则更新置信度分数

否则，

不更新；like

update the confidence score

otherwise,

not update;

其中Qⁱ、

分别表示一次优化后的第i个检测框的类别、位置坐标、置信度分数。6.6) Repeat steps 6.3)-6.5) until the preliminary test result set A is empty, and obtain an optimized test result set:

where Q ⁱ ,

Respectively represent the category, position coordinates, and confidence score of the ith detection frame after an optimization.

进行二次优化。Step 7: According to the structural characteristics of the ship in the optical remote sensing image, all the detection results in the detection result set B after one optimization are analyzed.

Perform secondary optimization.

的大小：若

则将第i个检测框的类别lⁱ优化为背景，否则优化为舰船目标；7.1) Judgment Confidence Score

size: if

Then the category li of the ^ith detection frame is optimized as the background, otherwise it is optimized as the ship target;

得到检测框的最长边x，判断x的大小：若x＞450，则将第i个检测框的类别Qⁱ优化为背景，否则，优化为舰船目标；7.2) According to the position coordinates of the detection frame

Obtain the longest side x of the detection frame, and judge the size of x: if x>450, optimize the category Q ⁱ of the i-th detection frame as the background, otherwise, optimize it as the ship target;

计算检测框的最长边与最短边之比y，判断y的大小：若y＞11，则将第i个检测框的类别Qⁱ优化为背景，否则，优化为舰船目标。7.3) According to the position coordinates of the detection frame

Calculate the ratio y of the longest side to the shortest side of the detection frame, and determine the size of y: if y>11, optimize the category Q ⁱ of the ith detection frame as the background, otherwise, optimize it as the ship target.

The effect of the present invention can be further illustrated by the following simulation experiments:

1. Simulation conditions:

The simulation experiment uses the full-color images of major Chinese cities captured by the "Gaofen-2" optical remote sensing satellite, with a ground resolution of 1 meter or 4 meters.

The CPU used in the simulation is Intel(R) Core(TM) i7-8750H, the main frequency is 2.20GHz, the GPU is GeForceGTX1080 with 8G, the simulation platform is the UBUNTU 16.04 operating system, the TensorFlow deep learning framework is used, and the Python3.6 software is used for experiments.

2. Simulation content and results:

Simulation 1, using the present invention to perform water and land segmentation on a 10000×10000 wide remote sensing image, and shield the land according to the results of the water and land segmentation. It can be seen from Figure 2 that the docked ships connected to the land are well preserved.

Simulation 2, the existing method based on convolutional neural network is used to directly detect the 10000×10000 wide remote sensing images one by one. Numbers are confidence time divisions. It can be seen from Fig. 3 that there is obvious missed detection due to the incomplete hull of the wide remote sensing image, which is because the distribution characteristics of ships are not considered in the existing methods.

Simulation 3, using the present invention to detect ship targets on a 10,000×10,000 wide-format remote sensing image, the result is shown in Figure 4, where the green box marks the detected ships, and the blue numbers are the confidence scores.

Comparing simulation 2 and simulation 3, the present invention effectively performs water and land segmentation on the premise of retaining the information of the ships docked at the shore, and extracts the region of interest according to the results of the water and land segmentation, and on this basis adopts a partial overlap detection strategy and improves non-maximum Compared with the existing methods, the detection efficiency is effectively improved, the detection accuracy of ship targets is improved, and the land false alarm is greatly reduced.

Claims (6)

1. a remote sensing image ship detection method based on block extraction of interest, is characterized in that, comprises as follows: (1) Constructing the optical remote sensing image ship detection dataset G: 1a) Download the optical remote sensing data of Gaofen-2, manually filter out the areas that contain ship targets, crop and save these areas in a partially overlapping manner; 1b) Randomly flip or rotate all the images obtained in 1a) up, down, left and right to obtain an amplified image and save it; 1c) Annotate the amplified image with a slanted rectangular frame, save the annotation information as an xml format file, and use all the amplified images and their corresponding annotation information to form the optical remote sensing image ship detection dataset G; (2) down-sampling the wide-format optical remote sensing image, and using the dehazing algorithm based on the dark channel prior to reduce the cloud and fog occlusion on the image, and obtain the enhanced image I; (3) Use the context information and image global features to perform water and land segmentation on the enhanced image I, extract the water area of interest, and obtain a binary image: (3a) using the context feature information of the pixel to perform preliminary threshold segmentation on the enhanced image I to obtain a preliminary water and land segmentation binary map I ₃ ; (3b) Mark the water connected domain set W _{= { w 1} , _. The set _{L = { l 1 ,} . The regional features of the domain and the land-connected domain are re-calibrated according to the discriminant rules, and the optimized binary map I ₄ of land and water segmentation of the enhanced image I is obtained; (3c) after performing the morphological expansion operation on the optimized water and land segmentation binary image I ₄ , the upsampling is restored to the original wide image size, and the final water and land segmentation binary image I ₅ is obtained; (4) Using the random multi-scale strategy to train the SCRDet target detection model M ₀ based on the convolutional neural network, and obtain the trained detection model M; the specific implementation is as follows: 4a) Take 90% of the constructed optical remote sensing image ship detection dataset G as training samples, and the remaining 10% as test samples; 4b) randomly select a training image and its corresponding label in the training sample, and randomly flip the image up and down or left and right and label it to obtain the transformed image X and the real label Y; 4c) Randomly select a certain scale from [600, 700, 832] as the short side length of the image, scale the transformed image X and the ground truth Y to the selected scale, and input it into the SCRDet model M ₀ to obtain the prediction result y'; 4d) Calculate the error between the predicted result Y' and the real label Y, and use the optimizer Adam to minimize the error to update the weight parameter of the SCRDet model M ₀ ; 4e) Repeat 4b)-4d), when the number of training rounds reaches 200,000, obtain the detection model M after training; (5) Using the final water and land segmentation binary image I ₅ , use a partially overlapping sliding window to extract the block of interest as the area to be detected on the original wide image, and form the image set of the area to be detected as F={f ₁ ,... ,f _i ,...,f _u }, the position set is

Wherein f _i represents the _ith detection area, _si represents the upper left corner coordinate of the detection area fi _, input the image set F of the area to be detected into the detection model M, and obtain the area detection result of each area fi; (6) Map the region detection results to the original wide image scale according to the position set S, and obtain a preliminary detection result set

where P ⁱ ,

Respectively represent the category, position coordinates, and confidence score of the ith detection frame of the preliminary detection; perform improved non-maximum suppression on the preliminary detection result set A, and obtain an optimized detection result set

where Q ⁱ ,

Respectively represent the category, position coordinates, and confidence score of the ith detection frame optimized once; (7) Perform secondary optimization on the primary optimization detection result set B according to the structural characteristics of the ship, and obtain the final detection result set

where R ⁱ ,

Represent the category, position coordinates, and confidence score of the finally generated i-th detection frame, respectively. 2. method according to claim 1, wherein utilize the context feature information of pixel to make preliminary threshold segmentation to enhanced image I in (3a), concrete realization is as follows: 3a1) Extend the edge of the enhanced image I with a mirror image with a width of 3 to obtain the extended image I', take each pixel of the enhanced image I as the center of a 7 × 7 sliding window, extract the corresponding window area on the extended image I' and Calculate the regional variance v as the context feature value of the central pixel point, and obtain the feature map V of the enhanced image I, and the feature value of the enhanced image I at the coordinates (i, j) is V(i, j); 3a2) Count the distribution histogram of the feature map V at intervals of 1, determine the peak variance v _max corresponding to the peak of the histogram, and calculate the valley value corresponding to the first valley value of the histogram in the range of [v _max , +∞) Variance g _t , if V(i, j) ≤ g _t , update V(i, j)=1, otherwise, update V(i, j)=0, and perform morphological expansion operation on the updated feature map V , obtain the characteristic binary image I ₁ ; 3a3) use the improved Laplacian operator to make grayscale transformation to the enhanced image I, utilize the OTSU adaptive threshold method to segment the enhanced image I, and obtain a grayscale binary image I ₂ ; 3a4) The feature binary image I ₁ and the grayscale binary image I ₂ are correspondingly performed pixel-by-pixel logical AND operation to obtain a preliminary water and land segmentation binary image I ₃ , where pixel value 0 in I ₃ represents land, and pixel value 1 represents water. 3. The method according to claim 1 , wherein in (3c), the optimized water and land segmentation binary graph I ₄ is used to obtain the final water and land segmentation binary graph 1 ₅ , which is realized as follows: 3c1) Copy the optimized binary image I ₄ of land and water segmentation, and obtain the copied result image I ₄ ′; 3c2) Use 15×15 ellipse structural elements to scan and optimize the binary image I ₄ of water and land segmentation pixel by pixel, and perform a logical AND operation between the structural element and each pixel in the corresponding area of I ₄ : 3c3) Judging the result of each pixel point, update the center pixel in the corresponding area of I ₄ ': If the result of each pixel is 0, update the center pixel of the corresponding area to 0; Otherwise, update the center pixel of the corresponding area to 1; 3c4) Upsampling the updated I ₄ ' by 8 times, restoring the size of I ₄ ' to the original wide image size, and obtaining the final water and land segmentation binary image I ₅ . 4. The method according to claim 1, wherein ( ₅ ) utilizes the final water and land segmentation binary image 15 to extract the block of interest as the region to be detected in the original wide-format image, and the region image set F to be detected is input into the The region detection result is obtained in the detection model M, and the specific implementation is as follows: 5a) Initialize the image set F of the area to be detected and the position set S to be empty sets; 5b) Scan the original wide remote sensing image with a partially overlapping window with a size of p × p and a step size of p/2, and calculate the final water and land segmentation binary map _I5 area corresponding to the window area f of the original wide remote sensing image. Water area, let the water area be S _w ; 5c) According to the size of S _w , determine whether the window area f is the block of interest: If S _w >(p/8) ² , then the window area f is determined to be the block of interest, the window area f is added to the image set F of the area to be detected, and the upper left corner coordinate s of the window area f is added to the position set S; Otherwise, it is determined that the window area f is not a block of interest, and no operation is performed; 5d) Input the image set F={f ₁ ,...,f _i ,...,f _u } into the detection model M to obtain the category and position coordinates of all detection frames of each area f _i , the confidence score. 5. The method according to claim 1, wherein in (6), the regional detection result is mapped on the original wide-width image scale, and the preliminary detection result set A is made to improve the non-maximum suppression, and the specific implementation is as follows: 6a) Assemble according to location

The coordinates _{s i of} the area _fi to be detected in the

where P ⁱ ,

Respectively represent the category, position coordinates, and confidence score of the i-th detection frame of the preliminary detection; 6b) Initialize the optimized detection result set B to be empty; 6c) Assume the categories of all detection frames in the preliminary detection result set A to be the same category, and set the index of the detection result with the highest confidence in A

Take the detection result from A according to the index r

After joining set B, and

remove from set A; 6d) Calculate the intersection ratio of each detection frame in the preliminary detection result set A and the detection frame with the highest confidence

6e) Judging the intersection ratio

The size relationship with the intersection and union ratio threshold t ₂ : like

update the confidence score

otherwise,

not update; 6f) Repeat the operations of steps 6c)-6e) until the preliminary detection result set A is empty, and obtain an optimized detection result set:

where Q ⁱ ,

Respectively represent the category, position coordinates, and confidence score of the ith detection frame after an optimization. 6. The method according to claim 1, wherein in (7), all detection results in the detection result set B after one optimization are performed according to the structural characteristics of the ship

The second optimization is carried out, and the specific implementation is as follows: 7a) Judgment Confidence Score

size: if

Then the category li of the ^ith detection frame is optimized as the background, otherwise it is optimized as the ship target; 7b) According to the position coordinates of the detection frame

The longest side of the detection frame is x, and the size of x is judged: if x>450, the category Q ⁱ of the ith detection frame is optimized as the background, otherwise, it is optimized as the ship target; 7c) According to the position coordinates of the detection frame

Calculate the ratio of the longest side to the shortest side of the detection frame as y, and determine the size of y: if y > 11, optimize the category Q ⁱ of the ith detection frame as the background, otherwise, optimize it as the ship target.

Images