CN103593458A - Mass image searching system based on color features and inverted indexes - Google Patents

Abstract

The invention provides a mass image searching system based on color features and inverted indexes. The mass image searching system is used in the following steps that firstly, a CIE1976L*a*b*c (Lab for short) color space with good uniformity is selected, K-means clustering is conducted on the space, and n types of colors are obtained; secondly, image pixels to be searched for are mapped according to the principle that color differences are minimum, and dimensionality reduced images are obtained; thirdly, the images obtained in the second step are divided into grids, main colors in grid units are obtained and used as representative colors, and each image is composed of a plurality of representative colors; fourthly, user-defined coding is conducted on the representative colors obtained in the third step, a piece of class text composed of a plurality of character codes is obtained finally, the class text is uploaded to an inverted index server, index building of mass images is completed, and therefore an image searching function can be completed.

Description

A Massive Image Retrieval System Based on Color Feature and Inverted Index

field of invention

The invention belongs to the technical field of pattern recognition and information processing, and relates to massive image processing on an e-commerce platform.

In particular, it relates to an implementation scheme of massive image retrieval based on color features and inverted index.

Background technique

The e-commerce service industry is experiencing its golden age of development. It is estimated that by 2015, the revenue of China's e-commerce service industry will exceed one trillion yuan, and then China will have the largest and most leading e-commerce service industry in the world. While e-commerce is booming, massive product images are also increasing at a geometric multiple. Therefore, how to quickly and effectively retrieve massive commodity images has become a new research trend. The product image content has obvious shape differences, for example, clothes and pants have great differences in shape. Color features are the most critical and commonly used features, but directly processing RGB colors of color images is quite time-consuming work, so how to reduce the complexity of color statistics is the first problem faced by massive image processing. The present invention proposes a method based on color space quantization and feature coding in order to quickly obtain image color features, and further extracts the main color information of the image by using the method of image gridding, and finally establishes a large number of image inverted indexes through feature coding.

Contents of the invention

The purpose of the present invention is to solve the problem of rapid image retrieval under the massive images formed by the rapid development of e-commerce, so that consumers can quickly retrieve the products they care about when faced with massive images. A fast and effective retrieval method for massive images on e-commerce platforms is provided.

In order to achieve the above object, the present invention adopts the following technical solutions:

A massive image retrieval system based on color features and inverted indexes, characterized in that it comprises the following steps:

Step 1: First, in order to solve the dimensionality disaster problem of computing color features in the RGB color space, and considering the uniformity of the color space, select the CIE1976L ^* a ^* b ^* color space with good uniformity, and use the K-means clustering method to CIE1976L ^* a ^* b ^* for clustering, clustering to 256 colors.

Step 2: Get all the images to be retrieved from the e-commerce platform, first convert the RGB color of the image into CIE1976L ^* a ^* b ^* color, and use the color of each pixel in the image according to the principle of minimum color difference and 256 obtained in step 1 The color is mapped, and the dimension of each pixel of the final image becomes 256 dimensions.

Step 3: Grid the image obtained in Step 2, and the grid size is 8*8. Count the main colors in each grid unit, and use each main color as the representative color of this grid unit. In the end each image will consist of 64 representative colors.

Step 4: Encode the 64 representative colors obtained in Step 3 through custom coding rules, and the last image will correspond to a class text composed of 64 character codes, upload such text to the inverted index server, Complete the index establishment of massive images, and then complete the image retrieval function.

The invention combines the fast and effective feature of text retrieval, and converts image features into text well. The problem of fast and effective retrieval of massive images is solved, and the present invention has the following advantages:

1. Starting from consumers' requirements for user experience, through the improvement of textualization of image features, fast and effective image retrieval effects can be completed;

2. From the perspective of the e-commerce platform, the present invention can effectively integrate massive image information in the platform through color features. Thereby providing users with a better consumption experience and bringing more website traffic.

3. From the perspective of information processing technology, the present invention combines the advantages of text retrieval very well, and grids the image, thereby retaining the contour information of the image part, which is good for commodity images that are sensitive to contour information. Effect.

Description of drawings

Accompanying drawing 1 is a frame diagram of the retrieval system;

Attachment 2 main color map;

Attached Figure 3 custom code;

Accompanying drawing 4 character code text;

Accompanying drawing 5 part experiment results.

Detailed ways

In order to make the purpose, technical solutions and beneficial effects of the present invention more clear, the present invention will be further described in detail below in conjunction with specific cases and with reference to the accompanying drawings.

The invention is a method for retrieving similar images among massive images on an e-commerce platform. This method can convert image features into text-like keyword features that can be indexed, so that the inverted index search engine can be used to complete the rapid retrieval of images. This retrieval method can well meet the needs of users for fast and effective retrieval methods, and at the same time, it can greatly increase the user experience of e-commerce platforms, and it has been well verified in practice that image retrieval and text retrieval, which are originally irrelevant The benefits of combining the two search methods.

Our test experiment hardware and software environment is:

Hardware environment:

Computer type: desktop;

CPU: Pentium(R) Dual-Core CPU E5600 2.93GHz

Memory: 4.00GB (3.49GB usable)

System type: 32-bit operating system

Graphics card: integrated graphics

Software Environment:

IDE: Visual Studio 2010

Image processing SDK: OpenCV2.3.1

Search engine: Apache Solr1.4.1

Development languages: C++, Python

As shown in Fig. 1, the present invention is to the retrieval flowchart of similar image, and the retrieval method to similar product image comprises the following steps:

Step 1: First, in order to be able to use image color features to establish an efficient index, the primary problem to be solved is the dimension of image color features. The dimension of RGB color space is 16777216 dimensions (256*258*256). , then retrieval based on color features becomes impractical. In order to solve the problem of the disaster of dimensionality and consider the uniformity of the color space, the present invention selects the CIE1976L ^* a ^* b ^* color space with good uniformity, and clusters the CIE1976L ^* a ^* b ^* with the K-means clustering method classes, clustered to 256 colors. Because these 256 colors do not contain gray-scale colors, the dimensionality reduction effect on black-and-white images or images with lighter colors is not good, so we cluster the color space to 248 dimensions, and cluster the gray-scale space to 8 dimensions, so that The color space of 256-dimensional color plus gray scale is obtained, which is called standard color space in the present invention, and each color therein is called standard color.

Step 2: Get all the images to be retrieved from the e-commerce platform for batch image feature extraction. In order to illustrate the process of image feature extraction more clearly, we use the feature extraction of an image as an example to describe. First, convert the RGB color of the image into CIE1976L ^* a ^* b ^* color; then calculate the color difference value between the color value of each pixel in the converted image and each color in the standard color space obtained in step 1 (this The invention selects the CIEDE1976 color difference formula, the color difference formula is as formula 1), calculates the color difference of the CIE1976L ^* a ^* b ^* color in the image and the standard color space, selects the standard color with the smallest color difference as the representative color of the pixel, and finally we get the image All 16777216-dimensional pixels in will be mapped to 256-dimensional standard colors. We call the image resulting from this process the mapped image.

${\mathrm{DE}}_{1976}(x_1,x_2)=\sqrt{({\left({\mathrm{DL}}^{*}\right)}^2+{\left({\mathrm{Da}}^{*}\right)}^2+{\left({\mathrm{DB}}^{*}\right)}^2)}$ (Formula 1)

in: $x_1={[{\mathrm{<figure-callout\; id="1"\; label="L"\; filenames="HDA0000419499120000021.png"\; state="\{\{state\}\}">L</figure-callout>}}_1^{*},a_1^{*},{\mathrm{<figure-callout\; id="1"\; label="b"\; filenames="HDA0000419499120000021.png"\; state="\{\{state\}\}">b</figure-callout>}}_1^{*}]}^T,$ $x_2={[L_2^{*},a_2^{*},b_2^{*}]}^T,$ $\mathrm{DL}={\mathrm{<figure-callout\; id="1"\; label="L"\; filenames="HDA0000419499120000021.png"\; state="\{\{state\}\}">L</figure-callout>}}_1^{*}-L_2^{*},$ $\mathrm{Da}=a_1^{*}-a_2^{*},$ $\mathrm{DB}={\mathrm{<figure-callout\; id="1"\; label="b"\; filenames="HDA0000419499120000021.png"\; state="\{\{state\}\}">b</figure-callout>}}_1^{*}-b_2^{*},$ DE ₁₉₇₆ is x ₁ , the color difference of two CIE1976L ^* a ^* b ^* colors.

Step 3: Grid the image obtained in Step 2. Suppose the image size is 200*200, the grid size is 8*8, and the size of each grid unit is 25*25. Count the color histogram in each grid unit, and get the color value with the largest ratio as the main color (ie representative color) of the grid unit. Finally, each image will be composed of 64 main colors, which we call the main color map , as shown in Figure 2 below.

Step 4: In order to better use the inverted search engine that is good at text retrieval, we encode the main color map obtained in step 3 through a custom encoding rule, that is, convert each main color into a four-letter character Character encoding, similar to keywords in text retrieval. The first two digits of the character code record the coordinate information of the main color, and the last two digits record the main color value. The conversion process is shown in Figure 3. The last image will correspond to a class text composed of 64 character codes, which is called character code text in the present invention. The character code text is shown in Figure 4. Upload the character code text and the ID corresponding to the image to the inverted index server to complete the index establishment of a large number of images, and then complete the image retrieval function.

Effect of the inventive method:

In order to verify the effect of the present invention, we have obtained a large amount of sports goods image data set in a certain domestic e-commerce platform. images. The image retrieval time is not higher than 14ms, which achieves a good real-time effect. Some experimental results of the present invention are shown in FIG. 5 .

Claims (1)

1. the massive image retrieval system based on color characteristic and inverted index, is characterized in that, comprises the steps:

Step 1: the CIE1976L that selects good uniformity ^*a ^*b ^*color space, and with K-means clustering method to CIE1976L ^*a ^*b ^*carry out cluster, cluster to 256 kind of color;

Step 2: obtain all images to be retrieved, first the RGB color of image is converted into CIE1976L ^*a ^*b ^*color, and 256 kinds of colors that each the pixel color in image is obtained according to aberration minimum principle and step 1 do and shine upon, the dimension of each pixel of final image becomes 256 dimensions;

Step 3: by the image lattice obtaining in step 2, sizing grid is n*n.Statistical color mass-tone in each grid cell, and using each mass-tone as the representative color of this grid cell, last every image will be comprised of n*n representative color;

Step 4: 64 representative colors that step 3 is obtained carry out character code by self-defining coding rule, the corresponding class text being formed by 64 character codes of last image meeting, this class text is uploaded to inverted index server, the index that completes large nuber of images is set up, and then can complete image retrieval function.

Images