TWI480809B - Image feature extraction method and device - Google Patents

Description

Image feature extraction method and device

The present application relates to the field of image processing technologies, and in particular, to an image feature extraction method and apparatus.

Image features include three aspects of color, texture, and shape. This application mainly deals with the shape characteristics of the image. The shape-based image feature extraction algorithm extracts and expresses the features in the image that reflect the shape of the object through mathematical means.

Image feature extraction is widely used. For example, in the field of image search, the search engine responsible for providing image search needs to compare the image in the image database with the image requested to be searched, thereby finding out that the image database is close to the image requested for search. image. In this type of search, the image features of the image requested for searching and the image features of the image in the database are compared. Therefore, image feature extraction of the image in advance becomes an indispensable step.

Regarding shape-based image feature extraction, the more common methods in the prior art are implemented based on Hough transform. The Hough transform is to map the points on the image plane to the lines on the parameter plane, and finally extract the image features through statistical properties. The principle can be described as follows: The equation of a straight line can be represented by y=k*x+b, where k and b are parameters, which are the slope and the intercept, respectively. The parameters of all the straight lines passing through a certain point (x ₀ , y ₀ ) satisfy the equation y ₀ =k*x ₀ +b. For the point where the brightness in the target image plane (x, y) satisfies the preset condition, use b=yk*x to find the corresponding line on the (k, b) plane, and assign all points on the line to 1 And when the n straight lines intersect at one point, the value of the point is assigned to the number of straight lines that pass. Then, for a straight line on the image plane, according to the above process, a family of straight lines can be obtained on the parameter plane (k, b), and the intersection of the straight lines on the parameter plane has the highest value, and the point can represent the target plane. A straight line, through which the straight line existing on the target plane can be detected through the point of the highest value on the statistical parameter plane.

Multiple lines, and so on. For calculating circles and arcs, it is similar.

In order to more clearly illustrate the above principle of the prior art, the following description is made with FIG. 1 as a target image. For the sake of simplicity, FIG. 1 is a picture of 10*10 pixels in size, in which a straight line exists, with the lower left corner of the picture as a coordinate At the origin, the line can be expressed as y=5. Let the background brightness of the picture be lower, and the brightness of the points on this line is higher. The method of detecting this line by Hough transform is as follows:

S1: detecting each point in FIG. 1 according to coordinates;

S2: when it is detected that the point on the target image (the coordinates of the point is (x ₀ , y ₀ )) is greater than a predetermined threshold, a b=y ₀ is identified on the parameter plane (shown in FIG. 2). a line of k*x ₀ , assigned a value of 1 for each point on the line of the logo (eg, named alpha);

S3: For the intersection point of the straight line identified on the parameter plane, the alpha value of the intersection point is set as the number of straight lines passing through the point. In fact, the alpha value of the intersection point can also be expressed as the sum of the alpha values of all the straight lines passing through the point at the point, and the alpha value of the intersection point set as described above is the number of straight lines passing the point. Expression is essentially the same truth.

After the above processing, for the straight line of y=5 in the target image shown in FIG. 1, on the parameter plane shown in FIG. 2, a point, that is, a point of k=0, b=5, and α of the point can be identified. If the value is the highest, the point (0, 5) on the parameter plane can represent the line y=5 on the target image. The (0,5) points on the parameter plane, ie 0 and 5, are exactly the slope and intercept of y=5 on the target plane, respectively, indicating the point identified on the parameter plane, and detecting the existence of the target plane. This line of y=5.

The above describes an example of a method for detecting a straight line in a target plane by the existing Hough transform method. For the case where there are multiple straight lines in the target plane, according to the above method, a plurality of highly assigned points can be obtained on the parameter plane, so that the points assigned by the high values on the parameter plane can represent multiple straight lines on the target plane.

For the detection of other shapes such as a circle, an arc, and the like in the target image, the principle is similar to the above process.

In the research and practice of the prior art, the inventors have found that the prior art has the following problems: the above feature extraction using the Hough transform method will inevitably involve floating-point operations, such as the case where the slope of the line has floating-point operations, of course. For more complex round and arc feature extraction, more floating point operations will be involved. Those skilled in the art know that floating-point operations impose high requirements on the computing power of CPUs and the like. Under the same hardware configuration, the existing Hough method involving floating-point operations has a lower calculation speed.

The purpose of embodiments of the present application is to provide an image feature extraction method and apparatus to improve the speed of image feature extraction.

In order to solve the above technical problem, an embodiment of the present application provides an image feature extraction method and apparatus, which are implemented by: an image feature extraction method, comprising: extracting an image of an object contained in an original image; The extracted image fills the boundary with a single color as the background, and makes the filled image the smallest square; the square image full image is scaled to the first predetermined size image, and the scaled image is segmented. a sub-image block of a second predetermined size that does not overlap each other; respectively calculate a luminance derivative of adjacent pixels in a horizontal, vertical, positive 45°, and negative 45° directions, and the number of derivative extreme points in each of the four directions And the total number of extreme points located on the four boundaries of the sub-image block as the feature vector of the sub-image block; the feature vector of all the sub-image blocks is used as the feature vector of the original image.

An image feature extraction device includes: a capture unit for extracting an image of an object contained in an original image; a filling unit that fills the extracted image with a single color as a background, and makes The filled image becomes a minimum square; the normalization processing unit includes a scaling unit and a dividing unit, and the scaling unit is configured to scale the square image full image to a first predetermined size image, and the dividing unit is used to zoom The subsequent image is divided into sub-image blocks of a second predetermined size that do not overlap each other; a luminance derivative calculation and statistical unit for respectively calculating luminance derivatives of adjacent pixels in the horizontal, vertical, positive 45°, and negative 45° directions And the number of derivative extreme points in the four directions and the total number of extreme points on the four boundaries of the sub-image block as the feature vector of the sub-image block; the synthesis unit, all the sub-pictures The feature vector of the block is used as the feature vector of the original image.

It can be seen from the technical solution provided by the above embodiments of the present application that an image of the contained object is extracted from the original image, and the extracted image is filled with a single color as a background, and the filled image is made. The minimum square is formed, and the square image full image is scaled to an image of a first predetermined size, and the scaled image is divided into sub-image blocks of a second predetermined size that do not overlap each other, and horizontal, vertical, and The luminance derivative of adjacent pixels in the positive 45° and negative 45° directions, the number of derivative extreme points in the four directions and the total number of extreme points on the four boundaries of the sub-image block are taken as the sub- The feature vector of the image block, the feature vector of all the sub-image blocks is used as the feature vector of the original image, and after the above processing, the shape feature of the image can be extracted, and since these processes only involve the shaping operation, the floating process is not involved. Point operation, therefore, under the same hardware configuration, the processing speed can be greatly improved compared to the prior art.

The embodiment of the present application provides an image feature extraction method and apparatus.

In order to enable a person skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described in the following with reference to the accompanying drawings in the embodiments of the present application. The described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present application without departing from the inventive scope should fall within the scope of the present invention.

As shown in FIG. 3, the flow of the image feature extraction method embodiment of the present application is as follows:

S310: Pull out an image of the contained object from the original image.

There are several methods for extracting the image of the object in which the original image is summarized. Some introductions are given below.

In the original image, there are often objects and backgrounds, and the background generally occupies the peripheral portion of the original image, and the object generally occupies the middle portion of the original image. Moreover, in the original image, there is a large difference in gray scale between the edge of the object and the pixels of the background. Therefore, this feature can be used to extract the image of the object in the original image, that is, according to the edge of the object and The grayscale difference that exists in the background extracts the image of the contained object from the original image.

A specific example of this step is given below. First look at finding the left and right boundaries of the region in which the object is located in the original image, including the following steps:

A1: Count the sum of the gray values on each column of all pixels of the original image.

For example, a 10*10 pixel image, each pixel has a gray value attribute. In this step, the sum of the gray values of all the pixels on each column is calculated as the gray value of the column pixels.

In order to facilitate the processing of computer software and hardware, the sum of the gray values of each column that can be counted is usually stored in the array.

A2: Calculate the difference between the gray values of the adjacent two columns of the original image from left to right, and record the abscissa x _{a of the} right column when the difference is greater than the threshold.

For example, scanning the gray values of the columns stored in the array in A1 from left to right, and sequentially calculating the difference between adjacent values in the array, and when scanning the difference is greater than the threshold, such as scanning to the second column and the The difference of the gray value of the three columns is 50, which is greater than the preset threshold value 30, and the abscissa x _{3 of} the third column is recorded, corresponding to the subscript of the third element in the array.

In this way, the left border in the original image indicating the location of the object is found.

It should be noted that when the scan difference is greater than the threshold, it may also be the abscissa of the left column when the difference is greater than the threshold, for example, the difference between the gray values of the second column and the third column is 50. , greater than the preset threshold 30, may also be the abscissa x _{2 of} the second column, corresponding to the subscript of the second element in the array. This method differs from the previously described column width of only one pixel and does not affect the overall effect of the method.

A3: Calculate the difference between the sum of the gray values of the adjacent two columns of the original image from right to left, and record the abscissa x _{b of the} left column when the difference is greater than the threshold.

This step is similar to A2, scanning and calculating the array, recording the subscript x _b , thus finding the right border in the original image indicating the location of the object.

The setting of the thresholds in A2 and A3 above may be set according to the empirical value. For example, when the value is greater than a certain value, the background and the object boundary may be significantly different. In this case, the value may be set as a threshold.

Looking at the upper and lower boundaries of the area where the object is located in the original image, the principle is similar to the above A1 to A3, including the following steps:

B1: Counts the sum of the gray values on each line of all pixels of the original image.

B2: Calculate the difference between the gray values of the adjacent two columns of the original image from top to bottom, and record the ordinate y _{a of the lower} row when the difference is greater than the threshold.

B3: the difference between the sum of the gray values of the adjacent two columns of the original image from bottom to top, and records the ordinate y _b of the upper row when the difference is greater than the result of the threshold.

Thus, the upper and lower boundaries in the original image indicating the position of the object are found, respectively y _a , y _b .

Then the image in the range of (x _a , x _b , y _a , y _b ) is the object image that is extracted from the original image.

In the above manner, the rectangular image in which the object is located in the original image is extracted, which is the easiest to implement. Of course, there are other slightly more complicated ways. For example, based on the above method, the calculation of the gray value difference values in the two diagonal directions is added to find the boundary of the object in the two diagonal directions. Then find out the octagonal image of the object in the original image. Of course, further increase of the direction, you can also get the image of the 16-sided, 32-sided shape of the object, no longer repeat them.

In addition, the original image may be divided into several sub-regions in the horizontal direction, and the left and right boundaries of the object on each sub-region are found by the foregoing manner; correspondingly, the original image is divided into several sub-regions in the vertical direction, using the foregoing The way to find the upper and lower boundaries of the object on each sub-area. In this way, you can get the polygon area where the object is located. Of course, this type of approach will be more complicated.

S320: Fill the boundary with a single color as a background, so that the image that is extracted becomes the smallest square.

The single color may be, for example, a color in which RGB is (0, 0, 0), and of course, other RGB colors. Generally, the color of filling RGB is (0, 0, 0), which is fast and does not constitute interference, thereby facilitating the calculation of the subsequent luminance derivative.

The boundary is filled with a single color to make the extracted image the smallest square, in order to facilitate segmentation of the image including the object into sub-image blocks of a predetermined size in a subsequent step. Figure 4 shows a schematic diagram of the image of the fill-out being the smallest square.

S330: The square image full image is scaled to an image of a first predetermined size, and the scaled image is divided into sub-image blocks of a second predetermined size that do not overlap each other.

The square image full image is scaled to an image of a first predetermined size, for example, the square image full image may be scaled to an image of 64*64 pixel size. Of course, it can also be an image that is scaled to a size of 128*128 pixels. Scale scaling refers to scaling the length and width in equal proportions.

The scaled image is divided into sub-image blocks of a second predetermined size that do not overlap each other, and may be an image of 16*16 pixels, or an image of 8*8 pixels, or 32*32 pixels. The size of the image.

The processing of this step is to normalize the square image including the object to standardize and simplify the subsequent processing. The first predetermined size and the second predetermined size may be preset, and the set size does not have a qualitative difference as long as it is within a reasonable range.

Hereinafter, the first predetermined size is 64*64 pixels, and the second predetermined size is 16*16 pixels as an example. When the second predetermined size takes 16*16 pixels, the image of the first predetermined size 64*64 pixels including the object will be divided into 4*4 block sub-image blocks.

In addition, it should be noted that the scaled image may also be divided into sub-image blocks of a second predetermined size with overlapping conditions, which may make the calculation process a bit cumbersome and may increase the characteristics of the final output. The dimension of the vector, but still a viable option.

S340: Calculate the luminance derivatives of adjacent pixels in the horizontal, vertical, positive 45°, and negative 45° directions, respectively, and the number of derivative extreme points in the four directions and the extreme values on the four boundaries of the sub-image block. The total number of points is used as the feature vector of the sub-image block.

The feature vector introduced here to indicate the characteristics of the image block is first introduced. It can be regarded as a five-element vector, that is, M(a, b, c, d, e). When processed according to S340, it will be initialized in the computer system that performs the processing, and then initialized. The five-element vector of the sub-image block is M(a, b, c, d, e) = M(0, 0, 0, 0, 0).

Next, the brightness derivative is introduced. The luminance derivative is defined as: luminance derivative = luminance difference / pixel pitch. The brightness value can generally be found using an algorithm for the human eye sensitivity curve. This algorithm is well known in the art. One solution is: brightness L = 116 / 3 * (0.212649 * R / 255 + 0.715169 * G / 255 + 0.072182 * B / 255), where RGB represents the color value. The brightness value is generally 1 for full light and 0 for dark. However, in processing, a floating point value of 0 to 1 is generally mapped to an integer range of 1 to 255. It can be seen that the brightness derivative can indicate the change in brightness between pixels. For extracting image shape features, generally, according to the attribute that the brightness difference between the edge or contour of the object and other parts of the image content is relatively obvious, the edge or contour of the object in the image is found, thereby The shape of the object in the image is described in mathematical form.

The above attributes can be described by the extreme values of the luminance derivative. Specifically, the luminance derivative extreme value expression of each adjacent pixel can be calculated one by one in a certain direction. When the luminance derivative extreme value of each adjacent pixel is calculated one by one in a certain direction, the luminance derivative before and after a certain pixel position changes symbolically, that is, the extreme value of the luminance derivative. Physically speaking, where the extreme points are calculated, it is likely to be the edge of the object and other parts of the image, or a feature in the image that indicates that a part of the object is different from the shape of the other part. It can be seen that these features are features that can indicate the shape characteristics of the object itself and, therefore, can be used as an amount representing the shape feature of the object.

For the calculation of the luminance derivatives of adjacent pixels in the horizontal, vertical, positive 45°, and negative 45° directions, an introduction to a specific calculation method is given below:

Calculate the luminance derivative of the sub-image block in the horizontal direction. If there is a luminance derivative extreme value and falls within the sub-image block, add 1 to b. If there are multiple extreme values, b is the extreme value. The number of a; if the extreme value exists on the sub-image block boundary, then a is increased by 1. It should be noted that the luminance derivative of the pixels of two adjacent columns can be calculated here.

Calculate the luminance derivative of the sub-image block in the vertical direction. If there is a luminance derivative extreme value and falls within the sub-image block, add 1 to c. If there are multiple extreme values, then c is the extreme value. Number; if the extreme value exists on the sub-image block boundary, a is incremented by 1. It should be noted that the luminance derivative of the pixels of two adjacent rows can be calculated here.

Calculate the luminance derivative of the sub-image block in the positive 45° direction. If there is a luminance derivative extreme value and falls within the sub-image block, add 1 to d. If there are multiple extreme values, then d is the pole. The number of values; a is incremented by 1 if the extreme value exists on the sub-image block boundary. It should be noted that the luminance derivative on the adjacent two columns of pixels is calculated here. It should be noted that the calculation here may be the luminance derivative on two adjacent pixels in the positive 45° direction.

Calculate the luminance derivative of the sub-image block in the negative 45° direction. If there is a luminance derivative extreme value and falls within the sub-image block, add 1 to e. If there are multiple extreme values, then e is the pole. The number of values; a is incremented by 1 if the extreme value exists on the sub-image block boundary. It should be noted that the calculation here may be the luminance derivative on two adjacent pixels in the negative 45° direction.

It can be seen that, after the above processing, in the five-element vector corresponding to the sub-image block, a is a case where the boundary of the object in the image block is located on the boundary of the image block. In addition to the case of a, that is, the case of dropping on the boundary of the sub-image block, b indicates the number of edges of the object inside the sub-image block having the shape feature in the horizontal direction, and c indicates the object inside the sub-image block. There are the number of edges with shape features in the vertical direction, d indicates the number of edges with the shape feature in the positive 45° direction of the object inside the sub-image block, and e indicates that the internal object of the sub-image block is at minus 45°. The presence of the number of edges with shape features in the direction.

The features derived from these above can be used to better represent the features of the shape of the object in the image.

Of course, the above statistical method is not the only way. Those skilled in the art should easily think of other feasible ways after seeing the above-mentioned manner, for example, statistical processing of a, b, c, d, e, It has a flexible way.

S350: The feature vector of all the sub-image blocks is used as the feature vector of the original image.

In the above S340, a five-element vector group of one sub-image block is obtained, and the five-element vector group can express the shape feature of the sub-image block.

Since the image is divided into a plurality of sub-image blocks that do not overlap each other in S330, in S350, the five-dimensional vector group of all the divided sub-image blocks is required to represent the shape feature of the image.

For example, the 64*64 pixel image processed by S320 is divided into 16*16 pixel non-overlapping blocks, and is divided into 4*4 blocks. A five-element vector group representing the shape features of each sub-image block, the 16 sub-image blocks are arranged to form 16 five-element vector groups, or a vector group of a total of 80 yuan. And this 80-element vector group can be used to indicate the shape characteristics of the image.

In addition, the following normalization processing can also be included:

S311: Compare the length and width of the image. If the length is longer than the width, rotate the image 90 degrees clockwise.

Rotation is to ensure that all images are in a basic shape. For example, a pen, the shape in the picture is placed vertically, there will be horizontal placement. In order to uniformly compare the shape of the pen in each picture, it is best to place the pictures in the same direction.

Alternatively, it may be rotated 90 degrees counterclockwise.

It should be noted that the object processed in S311 may be an image extracted from the original image in the step S310.

In addition, the following normalization processing can also be included:

S312: comparing the sum of the gray value of the upper half and the lower half of the extracted image or the minimum square image, if the sum of the gray values of the upper half is greater than the sum of the gray values of the lower half, The extracted image or the smallest square image is inverted.

Similar to the rotation above, this step is also to make the objects in the image have a uniform orientation. For example, an image shows an apple, and the apple is inverted. In most of the pictures showing Apple, Apple is in the process of being placed, so it is best to invert the image of the inverted Apple to compare it with other pictures. Obviously, for an object that is large and small, the sum of the gray values of the upper half of the image is greater than the sum of the gray values of the lower half; on the contrary, for objects that are small and large, In the image, the sum of the gray values of the lower half of the image is greater than the sum of the gray values of the upper half. The above processing involves only the shaping operation and does not involve floating-point operations. Therefore, under the same hardware configuration, the processing speed can be greatly improved compared with the prior art.

From the perspective of the actual search engine, according to the daily update of 1 million product graphics calculation, 12 pictures must be processed per second, and each picture is processed within 100ms. This is just a matter of averaging. Considering the peak of product updates for 4 hours per day, and the cost of accessing the disk and network, 50 images must be processed per second according to design requirements, and each image feature is completed within 20ms. If the Hough transform in the prior art is used, a straight line is recognized for a 200×200 pixel image, and a standard quad core server is used, which takes about 20 ms. Just the Hough line transformation itself is not enough time. If the circle is recognized, it takes longer.

However, in the embodiment of the present application, the method of processing the picture into blocks is faster because the floating point operation is not involved. Moreover, the processing of the image after the block can also make full use of the advantages of the existing multi-core processor. Also for a 200X200 picture, using the method in the embodiment of the present application, the entire process can be completed within 10ms.

An embodiment of an image feature extraction device in the present application is described below. FIG. 5 is a block diagram showing an embodiment of the device. As shown in FIG. 5, the device embodiment includes: a capture unit 51 for An image of the object contained in the image; a filling unit 52, filling the extracted image with a single color as a background, and making the filled image a minimum square; the normalization processing unit 53, including a scaling unit 531 and a dividing unit 532, the scaling unit 531 is configured to scale the square image full image to a first predetermined size image, and the dividing unit 532 is configured to divide the scaled image into two non-overlapping images. a sub-image block of a predetermined size; a luminance derivative calculation and statistics unit 54 for respectively calculating luminance derivatives of adjacent pixels in the horizontal, vertical, positive 45°, and negative 45° directions, and respectively exponential values in four directions The number of points, and the total number of extreme points on the four boundaries of the sub-image block as the feature vector of the sub-image block; the synthesizing unit 55, the feature vector of all the sub-image blocks as the original image Feature vector.

Preferably, in the device embodiment, the normalization processing unit 53 may further include:

The rotating unit 533 is configured to compare the length and width of the extracted image. If the length is greater than the width, the filled image is rotated 90 degrees clockwise. As shown in Figure 6.

Preferably, in the device embodiment, the normalization processing unit 53 may further include:

The inverting unit 534 is configured to compare the sum of the gray values of the upper half and the lower half of the extracted image, and if the sum of the gray values of the upper half image is greater than the sum of the gray values of the lower half, the filled graph Like an inversion. As shown in Figure 7.

Of course, the normalization processing unit 53 can also include the rotation unit 533 and the inversion unit 534 at the same time, as shown in FIG.

As described in the foregoing method embodiment, in the device embodiment, the capturing unit 51 extracts an image of the object contained in the original image, and specifically may have a large grayscale difference according to the edge of the object and the background. An image of the contained object is extracted from the original image.

Further, the capturing unit 51 extracts an image of the contained object from the original image according to a large gray scale difference between the edge of the object and the background, and includes: counting gray values of each column of all pixels of the original image. And; calculate the difference between the gray values of the adjacent two columns of the original image from left to right, and record the abscissa x _{a of the} right column when the difference is greater than the threshold; the adjacent two columns of gray from the right to the left of the original image The difference between the sum of the degrees, and record the abscissa x _{b of the} left column when the difference is greater than the threshold; count the sum of the gray values on each line of all pixels of the original image; calculate the adjacent two rows of the original image from top to bottom The difference between the gray values, and records the ordinate y _{a of the} lower row when the difference is greater than the threshold; the difference between the sum of the gray values of the adjacent two rows of the original image from the bottom to the top, and records the upper side when the difference is greater than the threshold The ordinate y _{b of the} line; then the image in the range of (x _a , x _b , y _a , y _b ) is the image of the object drawn from the original image.

The device embodiment can be located in a computer system, implemented by a combination of hardware, hardware, or hardware and software, and preferably, the device can be located in a computer system that implements search engine functionality.

It can be seen from the above embodiment that the image of the contained object is extracted from the original image, the extracted image is filled with a single color as the background, and the filled image becomes the smallest square, and the square image is The image is scaled to a first predetermined size image, and the scaled image is divided into sub-image blocks of a second predetermined size that do not overlap each other, and are respectively calculated horizontally, vertically, positively 45°, and negatively by 45°. The luminance derivative of the adjacent pixels in the direction, the number of the derivative extreme points in the four directions and the total number of extreme points located on the four boundaries of the sub-image block are used as the feature vectors of the sub-image block. The feature vector of all the sub-image blocks is used as the feature vector of the original image. After the above processing, the shape feature of the image can be extracted, and since these processes involve only the shaping operation, the floating point operation is not involved, and therefore, the same In the hardware configuration, the processing speed can be greatly improved compared to the prior art.

For the convenience of description, the above devices are described separately by function into various units. Of course, the functions of the various units may be implemented in the same or multiple software and/or hardware in the practice of the invention.

As can be seen from the description of the above embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of a software plus a necessary universal hardware platform. Based on such understanding, the technical solution of the present invention can be embodied in the form of a software product in essence or in the form of a software product, which can be stored in a storage medium such as a ROM/RAM or a disk. Optical disks, etc., include instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to perform the methods described in various embodiments of the present invention or portions of the embodiments.

The various embodiments in the specification are described in a progressive manner, and the same or similar parts between the various embodiments may be referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiment.

The invention is applicable to a wide variety of general purpose or special purpose computing system environments or configurations. For example: PCs, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, small Computers, large computers, decentralized computing environments including any of the above systems or devices, and more.

The invention may be described in the general context of computer-executable instructions executed by a computer, such as a program module. Generally, a program module includes routines, programs, objects, components, data structures, and the like that perform particular tasks or implement particular abstract data types. The invention may also be practiced in a distributed computing environment in which tasks are performed by remote processing devices that are connected through a communications network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

Although the present invention has been described by way of example, it is understood by those of ordinary skill in the art .

51. . . Capture unit

52. . . Filling unit

53. . . Normalized processing unit

531. . . Scaling unit

532. . . Split unit

54. . . Luminance derivative calculation and statistical unit

55. . . Synthetic unit

533. . . Rotating unit

534. . . Inverted unit

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings used in the embodiments or the description of the prior art will be briefly described below. Obviously, the drawings in the following description are merely Some of the embodiments described in the present application can also obtain other drawings based on these drawings without departing from the prior art for those skilled in the art.

1 is a target image in a prior art Hough transform; FIG. 2 is a parameter plane in a prior art Hough transform; FIG. 3 is a flowchart of an embodiment of the method of the present application; FIG. 4 is a pop-up in the embodiment of the present application. The image is a schematic diagram of a minimum square; FIG. 5 is a schematic diagram of the image of the filled-out image in the embodiment of the present application as a minimum square; FIG. 6 is a block diagram of the implementation of the device of the present application; Figure 8 is a block diagram of an implementation of the apparatus of the present application.

Claims (11)

An image feature extraction method includes: extracting an image of an object contained in an original image; filling the extracted image with a single color as a background, and making the filled image a minimum square; The square image full image is scaled to an image of a first predetermined size, and the scaled image is divided into sub-image blocks of a second predetermined size; respectively, the sub-image block is calculated horizontally, vertically, positively 45°, negative The luminance derivative of adjacent pixels in the 45° direction, the number of derivative extreme points in the four directions and the total number of extreme points on the four boundaries of the sub-image block as the characteristics of the sub-image block Vector; the feature vector of all sub-image blocks is used as the feature vector of the original image. The method of claim 1, wherein extracting an image of the contained object from the original image comprises: extracting the contained object from the original image according to a grayscale difference between the edge of the object and the background Image. The method of claim 2, wherein the image of the contained object is extracted from the original image according to the grayscale difference between the edge of the object and the background, including: counting gray of each pixel of the original image. The sum of the degrees; calculate the difference between the gray values of the adjacent two columns of the original image from left to right, and record the abscissa x _{a of the} right column when the difference is greater than the threshold; the adjacent two columns of the original image from right to left The difference between the sum of the gray values, and record the abscissa x _{b of the} left column when the difference is greater than the threshold; the sum of the gray values on each line of all pixels of the original image; calculate the original image from top to bottom The difference between the gray values of the column, and records the ordinate y _{a of the} lower row when the difference is greater than the threshold; the difference between the sum of the gray values of the adjacent two columns of the original image from bottom to top, and records the difference is greater than The threshold is the ordinate y _b of the upper row; then the image within the range of (x _a , x _b , y _a , y _b ) is the image containing the object that is extracted from the original image. The method of claim 1, wherein filling the extracted image with a single color as a background comprises: using the color of the extracted image as RGB (0, 0, 0) The background fills the border. The method of claim 1, wherein the dividing the scaled image into the second predetermined size sub-image block comprises: dividing the scaled image into a second predetermined size that does not overlap each other. Sub-image block. The method of claim 1, further comprising: comparing the length and width of the extracted image, and if the length is greater than the width, rotating the image by 90 degrees clockwise. The method of claim 1, further comprising: comparing the sum of the gray value of the upper half and the lower half of the extracted image or the minimum square image, if the sum of the gray values of the upper half is greater than The sum of the half-tone values is inverted, and the captured image or the smallest square image is inverted. An image feature extraction device includes: a capture unit for extracting an image of an object contained in an original image; a filling unit, filling the extracted image with a single color as a background, and filling the image The subsequent image becomes a minimum square; the normalization processing unit includes a scaling unit and a dividing unit, and the scaling unit is configured to scale the square image full image to an image of a first predetermined size, the dividing unit is configured to The scaled image is divided into sub-image blocks of a second predetermined size; a luminance derivative calculation and statistical unit for respectively calculating luminance derivatives of adjacent pixels in a horizontal, vertical, positive 45°, negative 45° direction, and The number of derivative extreme points in the four directions and the total number of extreme points on the four boundaries of the sub-image block as the feature vector of the sub-image block; the synthesis unit, which will be the sub-image block The feature vector is used as the feature vector of the original image. The device of claim 8, wherein the normalization processing unit further comprises: a rotation unit for comparing length and width of the image, if the length is greater than the width, the padded The image is rotated 90 degrees clockwise. The device of claim 8, wherein the normalization processing unit further comprises: an inverting unit for comparing the sum of the gray values of the upper half and the lower half of the image, if the upper half If the sum of the gray values of the image is greater than the sum of the gray values of the lower half, the filled image is inverted. The device of any one of claims 8 to 10, wherein the device is located in a computer system that implements a search engine function.