CN115511690A - Image processing method and image processing chip - Google Patents

Abstract

The invention relates to an image processing method and an image processing chip, which can adapt to images with various resolutions by converting the process of processing a large-resolution image once into the process of processing a small-resolution image for multiple times without limiting the resolution of the original image. Therefore, the wide image is divided into a plurality of narrow images, and in the calculation process of the Gaussian convolution of each narrow image, the width of each line of images stored in the middle calculation process is reduced, so that the hardware circuit area of an image processing chip can be greatly reduced, and the calculation accuracy of Gaussian convolution filtering can be ensured. In addition, the method ensures the continuity and the non-continuity of extreme point search at the boundary of different blocks by setting the multiplexing area, realizes the seamless butt joint of the extreme point search of adjacent blocks and ensures the continuity and the accuracy of data processing.

Description

Image processing method and image processing chip

technical field

The invention relates to the technical field of robot vision, in particular to an image processing method and an image processing chip.

Background technique

SIFT (Scale-Invariant Feature Transform, scale-invariant feature transformation) image scale space generation method is developed on the basis of image multi-scale space theory. SIFT algorithm is a robust image feature detection method with scale, Invariance to rotation, translation, scaling, etc. Widely used in video tracking, image 3D modeling, object recognition, image panorama stitching and other fields. Due to the large amount of calculation of the SIFT algorithm, in practical applications, with the improvement of the performance of the camera, the resolution of the image is getting higher and higher, the amount of information contained in each image is also increasing, and the data to be processed is also greatly increased. It is relatively difficult to implement image processing by purely using software, and it is difficult to meet real-time requirements. Therefore, many articles propose to use the high-speed parallel computing capability of the hardware circuit and adopt a high-speed parallel architecture design for the SIFT algorithm to meet the real-time requirements.

The SIFT algorithm needs to perform multiple Gaussian convolutions to establish a scale-space Gaussian pyramid. Gaussian convolution has the property of linear separability. A two-dimensional Gaussian filter function can be decomposed into the product of two one-dimensional Gaussian filter functions, that is, the cascade of row Gaussian filter and column Gaussian filter is used to achieve the purpose of filtering, which is beneficial to hardware circuits. Parallel architecture is implemented to save hardware resources. The SIFT algorithm scale space Gaussian pyramid is established in the process of hardware circuit realization, and factors such as computing speed, computing precision and hardware circuit area must be balanced. Since the Gaussian convolution of image data adopts a cascade method, the Gaussian convolution is performed by row first, and the Gaussian convolution result of the row needs to be temporarily stored, and then the Gaussian convolution is performed by column. The larger the length of the Gaussian convolution kernel, the higher the calculation accuracy, but the more image lines need to be stored in the middle of the calculation, resulting in a larger area of the hardware circuit. At the same time, the image width resolution is also a key factor affecting the area of the hardware circuit. The larger the image width resolution, the larger the area of the hardware circuit. In order to save the hardware circuit area, the existing technology usually limits the length of the Gaussian convolution kernel and the resolution of the image width, which has a great impact on the calculation accuracy of the SIFT algorithm and the application of large-resolution images. Therefore, how to improve the calculation accuracy and save the hardware circuit area has become a technical difficulty for the hardware circuit to realize the scale-space Gaussian pyramid of the SIFT algorithm.

Contents of the invention

In order to solve the above problems, the present invention provides an image processing method and an image processing chip, which can ensure higher calculation precision and accuracy while reducing the hardware circuit area. Concrete technical scheme of the present invention is as follows:

An image processing method, comprising the following steps: Step 1: Based on the resolution of the image to be processed, the image to be processed is divided into several blocks to be processed along the width direction of the resolution; Step 2: two adjacent The blocks to be processed are partially superimposed to form a multiplexing area; step 3: determine the length of the Gaussian convolution kernel, and perform Gaussian convolution filtering on the block to be processed to obtain a Gaussian pyramid image; step 4: based on Gaussian Pyramid image, the Gaussian difference pyramid image is obtained by subtracting the next layer image in the Gaussian pyramid image from the upper layer image; Step 5: Based on the Gaussian difference pyramid image, perform extreme value on the pixels of the block to be processed in the image to be processed Point search, the search range reaches the adjacent multiplexing area; Step 6: Obtain the stable feature points in the image to be processed according to the extreme point search results.

Further, the step 1 specifically includes the following steps: Step 11: Determine the width and height of the resolution of the image to be processed; Step 12: Take the height of the resolution of the image to be processed as the height unit, and take the preset resolution width as Width unit, divide the image to be processed along the width direction of the resolution to form a standard block to be processed; Step 13: Determine whether the width of the resolution of the undivided part of the image to be processed is smaller than the width unit , if no, return to step 12, continue to divide the standard block to be processed, if yes, directly use the undivided part as a non-standard block to be processed.

Further, the resolution width of the image to be processed in step 11 is 640pt, and the resolution height is 480pt; the preset resolution width described in step 12 is 84pt.

Further, the step 2 specifically includes the following steps: Step 21: Determine the start column boundary and the end column boundary of each block to be processed;

Step 22: Use the start column boundary of the current block to be processed as the multiplex start column boundary, and the end column boundary of the previous block to be processed as the multiplex end column boundary, the multiplex start column boundary and the multiplex The block between the end column boundaries is used as a multiplexing area, and the multiplexing area is provided with pixels of a preset number of columns.

Further, the preset number of columns in step 22 is 4 columns.

Further, the step 3 specifically includes the following steps: Step 31: Determine the length of the Gaussian convolution kernel as L, then the radius of the Gaussian convolution kernel is R, and R=(L-1)/2; Step 32: Search for the current waiting Process the current pixel of the block and the pixels within the radius of the Gaussian convolution kernel centered on the current pixel, and use the searched pixels to perform Gaussian convolution operations to obtain smoothed pixels; Step 33: After the pixels of the current block to be processed are smoothed, the pixels of the next block to be processed are smoothed, until the smoothing of the pixels of the entire image to be processed is completed, and a Gaussian image is obtained; step 34: Based on different Gaussian functions, repeat steps 32 and 33 to obtain multiple Gaussian images to form a set of Gaussian image groups; Step 35: Repeat steps 32, 33 and 34 based on images to be processed with different resolutions to obtain multiple Gaussian images Group of Gaussian images to form a Gaussian pyramid image.

Further, in the Gaussian pyramid image formed in step 35, the Gaussian image group at the bottom layer has 6 Gaussian images.

Further, the step 5 specifically includes the following steps: Step 51: Determine an image in the Gaussian difference pyramid image as the current image to be processed; Step 52: From the starting column of the current block to be processed in the current image to be processed The Nth pixel point counted backward from the boundary starts to search for the extreme value point, and the search ends at the Nth pixel point counted forward from the multiplexing end column boundary, and the extreme point point search of a row of pixels is completed; step 53 : Repeat step 52 until the extreme point search of each row of pixels in the current block to be processed is completed; Step 54: select the next block to be processed as the current block to be processed, and return to step 52 until the completion of the current block to be processed Extremum point search for pixels.

An image processing chip, comprising: a partition module, configured to divide the image to be processed into several blocks to be processed along the width direction of the resolution according to the resolution of the image to be processed, and two adjacent blocks to be processed The blocks are partially superimposed to form a multiplexing area; the operation module is used to perform Gaussian convolution filtering on the block to be processed according to the length of the Gaussian convolution kernel to obtain a Gaussian pyramid image, and based on the Gaussian pyramid image, The Gaussian difference pyramid image is obtained by subtracting the upper layer image from the next layer image in the Gaussian pyramid image; the analysis module is used to perform extreme value points on the pixels of the block to be processed in the image to be processed according to the Gaussian difference pyramid image Searching, the search range reaches the adjacent multiplexing area; the output module is used to output the stable feature points in the image to be processed according to the extreme point search results.

The effect of the present invention is that by converting the process of processing a large-resolution image into multiple processes of processing a small-resolution image, there is no limit to the resolution of the original image and it can adapt to images of various resolutions. In this way, an image with a large width is divided into multiple images with a small width. During the calculation process of the Gaussian convolution of each small-width image, the width of each line of image stored in the intermediate calculation process becomes smaller, which can greatly reduce the image. The hardware circuit area of the processing chip can also ensure the calculation accuracy of Gaussian convolution filtering. In addition, the method ensures the continuous and uninterrupted search of extreme points at the boundaries of different blocks by setting multiplexing areas, realizes the seamless connection of extreme point searches in adjacent blocks, and ensures the continuity of data processing and accuracy.

Description of drawings

FIG. 1 is a schematic flowchart of an image processing method according to an embodiment of the present invention.

FIG. 2 is a schematic diagram of dividing an image to be processed into blocks according to an embodiment of the present invention.

FIG. 3 is a schematic diagram 1 of overlapping blocks to be processed to form a multiplexing area according to an embodiment of the present invention.

FIG. 4 is a schematic diagram 2 of overlapping blocks to be processed to form a multiplexing area according to an embodiment of the present invention.

Fig. 5 is a schematic block diagram of the structure of the image processing chip according to the embodiment of the present invention.

detailed description

The technical solutions in the embodiments of the present invention will be described in detail below with reference to the drawings in the embodiments of the present invention. It should be understood that the specific embodiments described below are only used to explain the present invention, not to limit the present invention. Embodiments can be practiced by one of ordinary skill in the art without some of the specific details. For example, some circuits may be represented by circuit block diagrams, so as to avoid redundant and complicated descriptions of the embodiments in unnecessary details. Well-known circuits, structures and technical details may not be shown in detail so as not to obscure the embodiments.

The image processing method shown in FIG. 1 can be executed by a main control chip or a dedicated image processing chip in a vision robot. For the convenience of description, in the subsequent embodiments, the execution subject of the image processing method will be directly expressed as a robot. Described method specifically comprises the steps:

Step 1: First, the robot captures images through its own camera, and then the robot processes the captured images, and the images to be processed are called images to be processed. Based on the resolution of the image to be processed, the robot divides the image to be processed into several blocks to be processed along the width direction of the resolution. The resolution of the image to be processed is determined by the configuration of the image sensor of the robot, generally 640*480 pt, 1024*768 pt or 1920*1080 pt, etc. Among them, 640, 1024 and 1920 represent the width value of the image resolution, and 480, 768 and 1080 represent the height value of the image resolution. As shown in Figure 2, the outermost rectangular border represents the boundary of the image, the direction of the long side of the rectangle (that is, the horizontal line in the figure) is the width direction of the resolution, and the direction of the short side of the rectangle (that is, the vertical line in the figure) is the height direction of the resolution. Each small rectangular border marked with numbers such as 1, 2, and 3 represents a divided block to be processed. The image in Figure 2 is divided into n blocks to be processed. The value of n can be determined according to the specific image resolution. Set the size and the size of the block to be divided, and generally select a value greater than or equal to 6.

Step 2: The robot partially superimposes two adjacent blocks to be processed to form a reuse area. As shown in Figure 3, the outermost rectangular border in the figure is the boundary of an image to be processed, the large black dots in the figure represent pixels, and the continuous small black dots represent omissions. The value of n can be set according to the specific image resolution size and the size of the tiles to be divided. Assuming that n=2 is taken as an example, the pixels of the image to be processed are 480*360, then the image to be processed in Figure 3 is divided into 3 blocks to be processed, and the line segment between the two arrows marked by n-1 corresponds to The block is the first block to be processed, the block corresponding to the line segment between the two arrows marked by n is the second block to be processed, and the block corresponding to the line segment between the two arrows marked by n+1 It is the third block to be processed, and the pixel size of each block to be processed is 164*360, that is, each block to be processed includes 164 columns and 360 rows of pixels, wherein, between two adjacent blocks to be processed There are 4 pixels in between that are superimposed to form multiplexing area A and multiplexing area B.

Step 3: The robot determines the length of the Gaussian convolution kernel according to the built-in parameters of the system. When performing Gaussian convolution processing of images, the larger the length of the Gaussian convolution kernel, the higher the calculation accuracy, but the more image lines need to be stored in the middle of the calculation, resulting in a larger area of the hardware circuit. In order to save the hardware circuit area, the existing technology usually limits the length of the Gaussian convolution kernel, which will have a great impact on the calculation accuracy of the SIFT algorithm and the application of large-resolution images. Since the image processing method adopts the image block technology in the aforementioned steps, it can adopt a larger value of the Gaussian convolution kernel length under the premise of ensuring the accuracy of the operation. Then the robot performs Gaussian convolution filtering on the block to be processed according to the determined value of the Gaussian convolution kernel to obtain a Gaussian pyramid image. The construction method of the Gaussian pyramid image has been disclosed in the prior art, and will not be repeated here.

Step 4: Based on the Gaussian pyramid image, the robot subtracts the upper layer image from the next layer image in the Gaussian pyramid image to obtain a Gaussian difference pyramid image. The Gaussian difference pyramid, that is, the DOG (Difference of Gaussian) pyramid is built on the basis of the Gaussian pyramid. The first group and the first layer of the DOG pyramid are the first group and the second layer of the Gaussian pyramid minus the first group and the first layer. 1 layer is obtained. By analogy, each difference image is generated group by group and layer by layer, and all difference images form a difference pyramid image.

Step 5: Based on the Gaussian difference pyramid image, the robot performs extreme point search on the pixel points of the block to be processed in the image to be processed. When searching for extreme points, the robot needs to compare the upper and lower layers of the Gaussian difference pyramid with a total of 26 adjacent pixels around the same layer to determine the extreme points. When the robot searches for the extreme points of the pixels on the border of the blocks to be processed, the pixels of the adjacent blocks need to be used. If there is no multiplexing area between the adjacent blocks to be processed, and only If the pixel points of the current block to be processed are stored, the pixel points at the boundary of the current block to be processed cannot be searched for extreme points. In order to solve this problem, the image processing method described in this embodiment adopts the image redundancy segmentation method, and by superimposing several pixel points between two adjacent blocks to be processed for multiplexing, the extreme points can be When searching, the search starting point starts from the pixel point in the middle of the multiplexing area and ends at the pixel point in the middle of the next multiplexing area. As shown in Figure 3, when it is necessary to search for the extreme point of block n to be processed, start from the third pixel in the first row of multiplexing area A, and search from left to right to the second pixel in multiplexing area B. The pixel points end, and the extremum search of the first row of pixels of block n to be processed is completed. In the same way, the extremum search of the pixels in the second row, the third row to the last row is completed in sequence. When searching for extreme points on the leftmost and rightmost border pixels of the image to be processed, the values outside the borders can be ignored, and the processing results will not be greatly affected. Using the method of this embodiment, the continuous and uninterrupted search for extreme points can be performed on the boundaries of different blocks to be processed, realizing the seamless connection of the search for extreme points of adjacent blocks to be processed, and improving the continuity and accuracy of data processing .

Step 6: Since the determined extreme points have feature points in different fuzzy degrees and different scales, these feature points are the stable features to be extracted by the SIFT algorithm. Therefore, according to the search results of the extreme points, you can directly get The stable feature points in the image to be processed, these stable feature points can accurately reflect the information contained in the image to be processed.

The method converts a process of processing a large-resolution image into a process of processing a small-resolution image multiple times, has no limitation on the resolution of the original image, and can adapt to images of various resolutions. In this way, an image with a large width is divided into multiple images with a small width. During the calculation process of the Gaussian convolution of each small-width image, the width of each line of image stored in the intermediate calculation process becomes smaller, which can greatly reduce the image. The hardware circuit area of the processing chip can also ensure the calculation accuracy of Gaussian convolution filtering. In addition, the method ensures the continuous and uninterrupted search of extreme points at the boundaries of different blocks by setting multiplexing areas, realizes the seamless connection of extreme point searches in adjacent blocks, and ensures the continuity of data processing and accuracy.

As one of the implementation manners, the step 1 specifically includes the following steps: first, in step 11, the robot determines the width and height of the resolution of the image to be processed. Then enter step 12, the robot takes the height of the resolution of the image to be processed as the height unit and the width of the preset resolution as the width unit, and divides the image to be processed along the width direction of the resolution to form a standard block to be processed . Each standard pending tile has the same resolution width and height. Wherein, the preset resolution width can be correspondingly configured according to R&D and design requirements such as chip area and calculation accuracy, and generally can be configured as any value from 60pt to 90pt. Then enter step 13, the robot judges whether the width of the resolution of the undivided part in the image to be processed is smaller than the width unit, if not, it indicates that the undivided part can continue to be divided into standard pending blocks, and returns Step 12, continue to divide the standard block to be processed. If so, it indicates that the remaining part cannot meet the division requirement of a standard block to be processed, and the undivided part is directly regarded as a non-standard block to be processed. When the robot performs image processing, it treats standard blocks and non-standard blocks to be processed in the same way. The method described in this embodiment can be applied to images of different sizes and resolutions through the division of standard blocks to be processed, and as long as the resolution width and height of standard blocks to be processed are set reasonably, the maximum Improve the image processing rate.

As one of the implementation manners, the resolution width of the image to be processed in step 11 is 640pt, and the resolution height is 480pt. The preset resolution width described in step 12 is 84pt. Then, the image to be processed can be divided into 8 blocks to be processed, and 4 pixels overlap between two adjacent blocks to be processed. Through this division method, 8 standard tiles to be processed can be divided, and the best image processing effect can be achieved.

As one of the implementations, the step 2 specifically includes the following steps: first, in step 21, the robot determines the start column boundary and the end column boundary of each block to be processed, as shown in Figure 4, the boundary ai is The start column boundary of the first pending tile, boundary ck is the end column boundary of the first pending tile; boundary bj is the starting column boundary of the second pending tile, and boundary em is the second End column boundary of the tile to be processed; boundary dl is the start column boundary of the third tile to be processed, boundary go is the end column boundary of the third tile to be processed; boundary fn is the fourth tile to be processed The starting column boundary of , and the boundary hp is the ending column boundary of the fourth pending tile. Then enter step 22, use the start column boundary bj of the second block to be processed as the multiplexing start column boundary, and use the end point column boundary ck of the first block to be processed as the multiplexing end column boundary, thereby forming the first multiplex area bckj. By analogy, the multiplexing area deml and the multiplexing area fgon are formed. These multiplexing areas are provided with pixel points of a preset number of columns, which can be set correspondingly according to specific image processing requirements, and generally can be set to 4 to 8 columns. The method described in this embodiment can accurately define the scope of the multiplexing area by dividing the multiplexing area in the form of boundaries, and ensure the multiplexing efficiency of pixels in the multiplexing area.

As one of the implementation manners, the preset number of columns in step 22 is 4 columns. Based on the search range of extreme points, setting the pixels in the multiplexing area to 4 columns can achieve the maximum multiplexing efficiency and avoid unnecessary redundancy and waste of computing resources.

As one of the implementations, the step 3 specifically includes the following steps: first, in step 31, the robot determines that the length of the Gaussian convolution kernel is L according to the system built-in parameters, and then the radius of the Gaussian convolution kernel can be calculated as R, R=(L-1)/2. Then enter step 32, the robot searches the current pixel of the current block to be processed and the pixels within the radius of the Gaussian convolution kernel centered on the current pixel, and uses the searched pixels to perform Gaussian convolution operation to obtain smoothing after the pixels. The specific Gaussian convolution operation can refer to the existing operation method, and will not be repeated here. Then enter in step 33, after the pixel point of the current block to be processed is smoothed, carry out the smoothing process of the pixel point of the next block to be processed, until finishing the smoothing process of the pixel point of the whole image to be processed, obtain a Gaussian image. Then enter step 34, repeat step 32 and step 33 based on different Gaussian functions, obtain multiple Gaussian images, and form a set of Gaussian image groups. The Gaussian function can be preconfigured according to specific product design requirements. Finally, enter step 35, repeat step 32, step 33 and step 34 based on the images to be processed with different resolutions, to obtain multiple sets of Gaussian image groups to form a Gaussian pyramid image. The method for constructing a Gaussian pyramid has been disclosed in the prior art. The method of this embodiment differs from it mainly in steps 32 and 33. When the method of this embodiment performs Gaussian processing of an image, it proceeds in a manner of blocks to be processed one by one. In this way, images of different resolution sizes can be adapted. As for other method steps that are the same as those in the prior art, details will not be repeated here.

Assuming that the length of the Gaussian convolution kernel is selected as L=33pt, the radius of the Gaussian convolution kernel is R=(L-1)/2=16pt. Perform Gaussian convolution filtering on the image blocks to be processed with a resolution of 84*480, and build a Gaussian pyramid. Gaussian convolution can be used to have a linearly separable property, and a two-dimensional Gaussian filter function can be decomposed into two one-dimensional Gaussian filter functions. product. First, perform Gaussian row convolution filtering, and then, because the column Gaussian convolution filtering needs to use the row Gaussian convolution filtering results, the SRAM needs to store 33 rows of row filtering data, that is, the required pixel points for storage are 84*33 = 2772 indivual. If the image with the original resolution of 640*480 is not segmented, 640*33 = 21120 pixels need to be stored, and the storage area will be greatly increased. In order to reduce the area of the hardware circuit in the existing technology, the calculation accuracy is often sacrificed. The length of the Gaussian convolution kernel used is relatively small, generally 9, and the number of pixels to be stored is 640*9 = 5760, and the storage area is still relatively large. At the same time, the calculation accuracy of the Gaussian convolution filter will be greatly reduced, resulting in a greatly reduced stability of the feature points generated by the SIFT algorithm, which cannot meet the high-precision application requirements. The method described in this embodiment performs Gaussian convolution processing by partitioning, which can not only reduce the circuit area, but also ensure the calculation accuracy.

As one of the implementations, the Gaussian pyramid image formed in step 35 has 6 Gaussian images in the bottom Gaussian image group, which can improve the accuracy of data and ensure the quality of image processing.

As one of the implementation manners, the step 5 specifically includes the following steps: First, in step 51, the robot determines an image in the Gaussian difference pyramid image as the current image to be processed. Then enter step 52, the robot starts to search for the extreme value point from the Nth pixel point after the start column boundary of the current block to be processed in the current image to be processed, and searches to the end of the multiplexing column boundary forward The counted Nth pixel ends, and the extreme point search of a row of pixels is completed. The value of N can be set correspondingly according to the design requirements, and generally can be set to any value from 2 to 4, including 2 and 4. Then enter step 53, repeat step 52 until the extreme point search of each row of pixel points in the current block to be processed is completed. Finally enter step 54, select the next block to be processed as the current block to be processed, and then return to step 52 to search for the block until the extreme point search of the pixels of each block to be processed in the current image to be processed is completed .

Taking Figure 3 as an example, assuming that the image is an image in the Gaussian difference pyramid image, the robot searches for extreme points on the image. When the robot searches for the extreme point of block n to be processed, it starts to search from the third pixel point (counting from left to right) in the first row in multiplexing area A, and searches to the 3rd pixel point in multiplexing area B. 2 pixels (counting from left to right) end, in this way, the extreme point search of a row of pixels is completed. Similarly, the robot starts to search from the third pixel in the second row in multiplexing area A (counting from left to right), and searches to the second pixel in multiplexing area B (counting from left to right) ) ends, and in this way, the extreme point search of another row of pixels is completed. And so on, until the robot completes the extreme point search of the current block n to be processed. Similarly, after the robot completes the extreme point search for all the blocks to be processed, it completes the extreme point processing of an image in the Gaussian difference pyramid image. It should be noted that, since there is no multiplexing area on the outermost boundary of the first block to be processed and the last block to be processed, when searching for extreme points, the data outside the boundary can be directly ignored, and the search results are not affected. would have too much of an impact.

In the method of this embodiment, the SIFT algorithm adopts the image redundancy segmentation method when searching for extreme points, and the feature of this method is that every two adjacent blocks to be processed have 4 prime points to overlap, so in The boundaries of different blocks ensure the continuous and uninterrupted search of extreme points, realize the seamless connection of extreme point searches in adjacent blocks, and improve the accuracy of search data.

The image processing chip shown in FIG. 5 includes a partition module, an operation module, an analysis module and an output module connected in sequence. Wherein, the partitioning module is used to divide the image to be processed into several blocks to be processed along the width direction of the resolution according to the resolution of the image to be processed, and the part between two adjacent blocks to be processed superimposed to form a multiplexing area. The operation module is used to perform Gaussian convolution filtering on the block to be processed according to the length of the Gaussian convolution kernel to obtain a Gaussian pyramid image, and based on the Gaussian pyramid image, subtract the next layer image in the Gaussian pyramid image by A layer of image to get the difference of Gaussian pyramid image. The analysis module is used to search the pixel points of the block to be processed in the image to be processed according to the Gaussian difference pyramid image, and the search range reaches the adjacent multiplexing area. The output module is used to output the stable feature points in the image to be processed according to the extreme point search results. The chip described in this embodiment adopts the image redundancy segmentation method, uses the partition module to divide the image with large resolution into multiple small-resolution images, and very cleverly converts the process of processing a large-resolution image into multiple The process of processing small resolution images for the first time. This chip has no limitation on the resolution of the original image, and can adapt to images of various resolutions. When establishing a scale-space Gaussian pyramid, it is processed according to the standardized small block image, which has nothing to do with the size and resolution of the original image, and supports greater resolution. rate and match range. In the standardized small-block image Gaussian convolution filtering process, the width of each line of image stored in the intermediate calculation process is very small, which can greatly reduce the area of the hardware circuit, which is conducive to the use of a larger Gaussian convolution kernel length to improve the accuracy of the SIFT algorithm . The image redundancy segmentation method is used in the establishment of the scale-space Gaussian pyramid, which has high application value for the realization of the high real-time and high-precision hardware circuit of the SIFT algorithm.

Those skilled in the art can understand that all or part of the steps in the method of the above-mentioned embodiments can be completed by instructing the relevant hardware through a program. (processor) Execute all or part of the steps of the methods described in the various embodiments of the present invention. These storage media can include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes. .

In addition, it should be noted that the various specific technical features described in the above specific implementation manners may be combined in any suitable manner if there is no contradiction. In order to avoid unnecessary repetition, various possible combinations are not further described in the embodiments of the present invention.

Claims (9)

1. An image processing method, characterized by comprising the steps of:

step 1: dividing an image to be processed into a plurality of image blocks to be processed along the width direction of the resolution based on the size of the resolution of the image to be processed;

step 2: part of two adjacent image blocks to be processed are overlapped to form a multiplexing area;

and step 3: determining the length of a Gaussian convolution kernel, and performing Gaussian convolution filtering on the image block to be processed to obtain a Gaussian pyramid image;

and 4, step 4: based on the Gaussian pyramid images, subtracting the next layer image from the previous layer image in the Gaussian pyramid images to obtain Gaussian difference pyramid images;

and 5: based on the Gaussian difference pyramid image, searching the extreme points of pixel points of blocks to be processed in the image to be processed, wherein the searching range reaches the adjacent multiplexing area;

step 6: and obtaining stable characteristic points in the image to be processed according to the extreme point search result.

2. The method according to claim 1, characterized in that said step 1 comprises in particular the steps of:

step 11: determining the width and height of the resolution of the image to be processed;

step 12: dividing the image to be processed along the width direction of the resolution by taking the height of the resolution of the image to be processed as a height unit and the width of a preset resolution as a width unit to form a standard image block to be processed;

step 13: and judging whether the width of the resolution of the part which is not divided in the image to be processed is smaller than the width unit, if not, returning to the step 12, continuing to divide the standard image block to be processed, and if so, directly taking the part which is not divided as a non-standard image block to be processed.

3. The method of claim 2, wherein:

the width of the resolution of the image to be processed in the step 11 is 640pt, and the height of the resolution is 480pt;

the preset resolution width in the step 12 is 84pt.

4. The method according to claim 1, wherein the step 2 comprises the following steps:

step 21: determining a starting column boundary and an end column boundary of each block to be processed;

step 22: the method comprises the steps of taking a starting column boundary of a current block to be processed as a multiplexing starting column boundary, taking an ending column boundary of a block to be processed as a multiplexing ending column boundary, taking a block between the multiplexing starting column boundary and the multiplexing ending column boundary as a multiplexing area, and setting a preset number of columns of pixel points in the multiplexing area.

5. The method of claim 4, wherein:

the number of the preset columns in the step 22 is 4.

6. The method according to claim 1, characterized in that said step 3 comprises in particular the steps of:

step 31: determining that the length of the Gaussian convolution kernel is L, and then the radius of the Gaussian convolution kernel is R, wherein R = (L-1)/2;

step 32: searching a current pixel point of a current block to be processed and a pixel point within the radius range of a Gaussian convolution kernel with the current pixel point as a center, and performing Gaussian convolution operation by using the searched pixel point to obtain a pixel point after smoothing processing;

step 33: after all the pixels of the current block to be processed are subjected to smoothing processing, smoothing processing is carried out on the pixels of the next block to be processed until smoothing processing of the pixels of the whole image to be processed is completed, and a Gaussian image is obtained;

step 34: repeating the step 32 and the step 33 based on different Gaussian functions to obtain a plurality of Gaussian images and form a group of Gaussian image groups;

step 35: and (4) repeating the steps 32, 33 and 34 based on the images to be processed with different resolutions to obtain a plurality of groups of Gaussian image groups to form Gaussian pyramid images.

7. The method of claim 6, wherein:

the gaussian pyramid image formed in step 35 has 6 gaussian images in the lowest gaussian image group.

8. The method according to claim 4, wherein the step 5 comprises the following steps:

step 51: determining one image in the Gaussian difference pyramid images as a current image to be processed;

step 52: searching an extreme point from an Nth pixel point counted from the initial column boundary of the current block to be processed in the current image to be processed backward, and finishing searching until the Nth pixel point counted from the multiplexing ending column boundary, thereby finishing searching the extreme points of a row of pixel points;

step 53: repeating the step 52 until the searching of the extreme points of the pixel points of each row in the current block to be processed is completed;

step 54: and selecting the next block to be processed as the current block to be processed, and returning to the step 52 until the searching of the extreme points of the pixel points in the current image to be processed is completed.

9. An image processing chip, comprising:

the partitioning module is used for dividing the image to be processed into a plurality of image blocks to be processed along the width direction of the resolution according to the resolution of the image to be processed, and partially overlapping two adjacent image blocks to be processed to form a multiplexing area;

the operation module is used for carrying out Gaussian convolution filtering on the image blocks to be processed according to the length of the Gaussian convolution kernel to obtain Gaussian pyramid images, and subtracting the upper layer of image from the lower layer of image in the Gaussian pyramid images to obtain Gaussian difference pyramid images based on the Gaussian pyramid images;

the analysis module is used for searching extreme points of pixel points of blocks to be processed in the images to be processed according to the Gaussian difference pyramid images, and the searching range reaches an adjacent multiplexing area;

and the output module is used for outputting the stable characteristic points in the image to be processed according to the extreme point search result.

Images