Approximation calculation-based vgg image feature extraction acceleration method
By selecting a cyclic procedure that matches the target image and generating an approximation procedure in the VGG network, the problems of computational complexity and long processing time of the VGG network are solved, thereby accelerating and improving the efficiency of image feature extraction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUIZHIAN INFORMATION TECH CO LTD
- Filing Date
- 2022-07-08
- Publication Date
- 2026-06-09
AI Technical Summary
Existing VGG networks are computationally complex, time-consuming, and memory-intensive in image recognition, especially in large-scale image recognition scenarios where they are inefficient and cannot achieve timely and accurate feature extraction.
By selecting a target loop program that matches the target image from the set of image feature extraction programs of the VGG network, executing the initial approximation program for piercing generation, and determining whether to perform secondary piercing based on the image feature recognition results, the final approximation program is generated, thereby reducing the computational workload of subsequent target loop programs.
It accelerates image feature extraction, reduces the computational workload of subsequent target loop programs, and improves the efficiency and accuracy of image feature recognition.
Smart Images
Figure CN115222971B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image recognition and processing, and in particular to an accelerated method for VGG image feature extraction based on approximate calculation. Background Technology
[0002] VGG networks are commonly used image recognition and analysis networks, comprising several different types of image recognition programs. In practical applications, an appropriate image recognition program can be selected based on the specific characteristics of the image to perform matching-based image feature extraction. However, VGG network-based image recognition programs involve complex computation processes, long computation times, and significant memory consumption. Especially in image feature recognition scenarios with a large number of images, they cannot quickly analyze all images, reducing the efficiency of image recognition and analysis and hindering timely and accurate feature extraction. Summary of the Invention
[0003] To address the shortcomings of existing technologies, this invention provides an accelerated VGG image feature extraction method based on approximation computation. First, a target loop program matching the target image is selected from a set of image feature extraction programs based on VGG networks. This target loop program is then perforated to obtain an initial approximation program. This initial approximation program possesses image recognition capabilities consistent with the target loop program, and it also boasts a higher computational speed compared to the target loop program. Next, based on the difference between the image feature recognition results of the target loop program and the initial approximation program for the target image, a secondary perforation of the initial approximation program is determined. Finally, the final approximation program's image feature recognition results for the image to be recognized are input into the target loop program, thereby obtaining the final image feature information for each image to be recognized. This method generates an initial approximation program through perforation for preliminary feature recognition of the image to be recognized, thus reducing the computational workload of the subsequent target loop program and accelerating image feature extraction.
[0004] This invention provides an accelerated method for VGG image feature extraction based on approximate calculation, which includes the following steps:
[0005] Step S1: Obtain the target image uploaded by the client terminal, perform preliminary screening on the target image, and determine the image type information of the target image; based on the image type information, select a target loop program that matches the target image from the set of image feature extraction programs based on VGG network;
[0006] Step S2: Perform perforation on the selected target loop program to obtain an initial approximation program; input the target image into the target loop program and the initial approximation program respectively for analysis and processing, thereby extracting the first image feature information and the second image feature information;
[0007] Step S3: Based on the difference in image feature recognition accuracy between the first image feature information and the second image feature information, determine whether to perform secondary perforation on the initial approximation procedure; and based on the above determination result, obtain the final form of the approximation procedure;
[0008] Step S4: Input several images to be identified that have the same image type information as the target image into the approximation program of the final form in sequence, and input the image feature recognition result corresponding to each image to be identified into the target loop program, so as to obtain the final image feature information of each image to be identified.
[0009] Furthermore, in step S1, acquiring the target image uploaded by the client terminal and performing preliminary screening processing on the target image to determine the image type information of the target image specifically includes:
[0010] The target image uploaded by the client terminal is acquired, and the target image is subjected to image brightness analysis and image sharpness analysis to obtain the image brightness value and image pixel resolution value of the target image.
[0011] If the image brightness value is greater than or equal to a preset brightness threshold and the image pixel resolution value is greater than or equal to a preset resolution threshold, then the target image is determined to be a high-quality image; otherwise, the target image is determined to be a low-quality image.
[0012] Further, in step S1, selecting a target loop program matching the target image from the set of image feature extraction programs based on the VGG network according to the image type information specifically includes:
[0013] When the target image is a high-quality image, a target loop program that matches the image color distribution features and the image contour distribution features is selected from the set of image feature extraction programs based on VGG networks, according to the image color distribution features and the image contour distribution features of the target image.
[0014] Furthermore, in step S2, the selected target loop program is perforated to obtain an initial approximate program, specifically including:
[0015] The target loop program selected above is subjected to a predetermined number of continuous loop punching operations to obtain an initial approximate program.
[0016] Further, in step S2, the target image is input into the target loop program and the initial approximation program for analysis and processing, thereby extracting the first image feature information and the second image feature information. Specifically, this includes:
[0017] The target image is sequentially pixel-sharpened and then input into the target loop program and the initial approximation program for analysis and processing, thereby extracting the first image feature information and the second image feature information respectively; wherein, the first image feature information and the second image feature information refer to the edge line contour features of people or objects in the target image.
[0018] Furthermore, in step S3, determining whether to perform a secondary perforation on the initial approximation procedure based on the difference in image feature recognition accuracy between the first image feature information and the second image feature information specifically includes:
[0019] Determine the degree of overlap between the first image feature information and the second image feature information in terms of the line contours at the edge lines;
[0020] If the overlap value of the thread profile is greater than or equal to the prediction threshold, it is determined that no secondary perforation is required for the initial approximation procedure; otherwise, it is determined that secondary perforation is required for the initial approximation procedure.
[0021] Furthermore, in step S3, based on the above judgment result, the approximate procedure for obtaining the final form specifically includes:
[0022] If it is determined that a second perforation is not required for the initial approximation procedure, then the current initial approximation procedure is used as the final form of the approximation procedure.
[0023] When it is determined that a second perforation is required for the initial approximation procedure, the initial approximation procedure after at least one round of second perforation is taken as the final form of the approximation procedure, provided that the final value of the thread wheel contour overlap is greater than or equal to the prediction degree threshold.
[0024] Further, in step S4, several images to be identified that have the same image type information as the target image are sequentially input into the final form of the approximation program, and the image feature recognition result corresponding to each image to be identified is input into the target loop program, thereby obtaining the final image feature information of each image to be identified, specifically including:
[0025] Acquire several images that belong to the same high-quality image category as the target image and have the same image color distribution characteristics and image contour distribution characteristics; and input each image to be identified sequentially into the final form of the approximation program to obtain the corresponding image feature recognition result;
[0026] The image feature recognition result of each image is then input into the target loop program to obtain the final image feature information of each image to be recognized, and to construct a one-to-one mapping relationship between each image to be recognized and its corresponding final image feature information.
[0027] Compared to existing technologies, this VGG image feature extraction acceleration method based on approximation computation first selects a target loop program matching the target image from a set of image feature extraction programs based on VGG networks, and then performs perforation on the target loop program to obtain an initial approximation program. This initial approximation program has image recognition capabilities consistent with the target loop program, and it also has a higher computational speed. Next, based on the difference between the image feature recognition results of the target loop program and the initial approximation program, it is determined whether to perform secondary perforation on the initial approximation program. Finally, the final form of the approximation program is input into the target loop program to obtain the final image feature information for each image to be recognized. This method generates an initial approximation program through perforation for preliminary feature recognition of the image to be recognized, thus reducing the computational workload of the subsequent target loop program and accelerating image feature extraction.
[0028] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.
[0029] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0031] Figure 1 This is a flowchart illustrating the accelerated VGG image feature extraction method based on approximate calculation provided by the present invention. Detailed Implementation
[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0033] See Figure 1 This is a schematic flowchart illustrating the accelerated VGG image feature extraction method based on approximate calculation provided in an embodiment of the present invention. The accelerated VGG image feature extraction method based on approximate calculation includes the following steps:
[0034] Step S1: Obtain the target image uploaded by the client terminal, perform preliminary screening on the target image, and determine the image type information of the target image; based on the image type information, select the target loop program that matches the target image from the set of image feature extraction programs based on VGG network;
[0035] Step S2: Perform perforation on the selected target loop program to obtain an initial approximation program; input the target image into the target loop program and the initial approximation program respectively for analysis and processing, thereby extracting the first image feature information and the second image feature information;
[0036] Step S3: Based on the difference in image feature recognition accuracy between the first image feature information and the second image feature information, determine whether to perform secondary perforation on the initial approximation procedure; and based on the above determination result, obtain the final form of the approximation procedure;
[0037] Step S4: Several images to be identified that have the same image type information as the target image are sequentially input into the final form of the approximation program, and the image feature recognition result corresponding to each image to be identified is input into the target loop program, thereby obtaining the final image feature information of each image to be identified.
[0038] The beneficial effects of the above technical solution are as follows: This VGG image feature extraction acceleration method based on approximation calculation first selects a target loop program matching the target image from the set of image feature extraction programs based on VGG networks, and performs perforation on the target loop program to obtain an initial approximation program. In this way, the initial approximation program can have the same image recognition function as the target loop program, and the initial approximation program has a higher computational speed than the target loop program. Then, based on the difference between the image feature recognition results of the target image by the target loop program and the initial approximation program, it is determined whether to perform secondary perforation on the initial approximation program. Finally, the image feature recognition result of the final approximation program is input into the target loop program to obtain the final image feature information of each image to be recognized. The above method generates an initial approximation program in a perforated manner for preliminary feature recognition of the image to be recognized, which can reduce the computational workload of the subsequent target loop program and accelerate image feature extraction.
[0039] Preferably, in step S1, acquiring the target image uploaded by the client terminal, performing preliminary screening on the target image, and determining the image type information of the target image specifically includes:
[0040] The system acquires the target image uploaded by the client terminal and performs image brightness analysis and image sharpness analysis on the target image to obtain the image brightness value and image pixel resolution value of the target image.
[0041] If the image brightness value is greater than or equal to a preset brightness threshold and the image pixel resolution value is greater than or equal to a preset resolution threshold, then the target image is determined to be a high-quality image; otherwise, the target image is determined to be a low-quality image.
[0042] The beneficial effects of the above technical solution are as follows: The image feature extraction program set based on VGG networks includes multiple different types of target loop programs, each of which can achieve high image feature recognition performance for specific types of images. To ensure high-precision image feature recognition, the target image is first preliminarily screened based on its brightness and pixel resolution to determine whether it is a high-quality image. Only when the target image is of high quality can a suitable target loop program be selected based on it. Furthermore, when the target image is of low quality, pixel interpolation and brightness adjustment can be performed first to meet the standards of a high-quality image before proceeding with subsequent operations.
[0043] Preferably, in step S1, selecting a target loop program matching the target image from the set of image feature extraction programs based on the VGG network according to the image type information specifically includes:
[0044] When the target image is a high-quality image, a target loop program that matches the image's color distribution characteristics and contour distribution characteristics is selected from the set of image feature extraction programs based on the VGG network.
[0045] The beneficial effects of the above technical solution are as follows: By using the above method, based on the color distribution features and contour distribution features of the target image, a matching target loop program is selected from the set of image feature extraction programs based on VGG network. This ensures that the selected target loop program can adapt to its color and contour to perform optimized and accurate feature recognition, thereby improving the matching between the target image and the target loop program.
[0046] Preferably, in step S2, the selected target loop program is perforated to obtain an initial approximate program, specifically including:
[0047] The target loop program selected above is subjected to a predetermined number of continuous loop punching operations to obtain an initial approximate program.
[0048] The beneficial effects of the above technical solution are as follows: by using the continuous loop punching method, the target loop program is approximated to obtain the corresponding initial myopia program. This ensures that the initial approximation program and the target loop program have the same image feature recognition function. Moreover, the initial approximation program can quickly perform image feature recognition with less computation and less memory usage compared to the target loop program. It also makes it easier to use the initial approximation program as a pre-processing program for the target loop program, thereby reducing the computational load of the target loop program.
[0049] Preferably, in step S2, the target image is input into the target loop program and the initial approximation program for analysis and processing, thereby extracting the first image feature information and the second image feature information, specifically including:
[0050] After the target image is processed by pixel sharpening, it is input into the target loop program and the initial approximation program for analysis and processing, thereby extracting the first image feature information and the second image feature information respectively; wherein, the first image feature information and the second image feature information refer to the edge line contour features of people or objects in the target image.
[0051] The beneficial effects of the above technical solution are as follows: through the above method, the quantitative results of feature recognition of the target image by the target loop program and the initial approximation program can be obtained, which facilitates the subsequent judgment on whether the initial approximation program has achieved the corresponding image feature recognition accuracy.
[0052] Preferably, in step S3, determining whether to perform a secondary perforation on the initial approximation procedure based on the difference in image feature recognition accuracy between the first image feature information and the second image feature information specifically includes:
[0053] Determine the degree of overlap between the first image feature information and the second image feature information in terms of the line contours at the edge lines;
[0054] If the overlap value of the thread profile is greater than or equal to the prediction threshold, it is determined that no secondary piercing is required for the initial approximation procedure; otherwise, it is determined that secondary piercing is required for the initial approximation procedure.
[0055] The beneficial effects of the above technical solution are as follows: By using this method, it is possible to quantitatively determine whether the initial approximation program has reached the corresponding image feature recognition accuracy, thereby deciding whether to perform secondary piercing on the initial approximation program more than once. In fact, each time the initial approximation program undergoes a round of secondary piercing, the closer it gets to the target loop program (i.e., the closer the image feature recognition accuracy is to the target loop program, the longer the image feature recognition time and the larger the memory space usage), the better. In actual operation, it is desirable to perform secondary piercing on the initial approximation program as few times as possible, but it is also necessary to set an appropriate number of secondary piercing operations on the initial approximation program based on the results of the above threshold comparison.
[0056] Preferably, in step S3, the approximate procedure for obtaining the final form based on the above judgment result specifically includes:
[0057] If it is determined that no secondary perforation is required for the initial approximation procedure, then the current initial approximation procedure is used as the final form of the approximation procedure.
[0058] When it is determined that a second perforation is required for the initial approximation procedure, the initial approximation procedure after at least one round of second perforation is taken as the final form of the approximation procedure, provided that the final value of the thread wheel profile overlap is greater than or equal to the prediction degree threshold.
[0059] The beneficial effects of the above technical solution are as follows: by using the above method, it is possible to ensure that the final approximation program achieves the optimal balance between image feature recognition accuracy and image feature recognition time consumption and memory space occupancy.
[0060] Preferably, in step S4, several images to be identified that have the same image type information as the target image are sequentially input into the final form of the approximation program, and the image feature recognition result corresponding to each image to be identified is input into the target loop program, thereby obtaining the final image feature information of each image to be identified, specifically including:
[0061] Acquire several images that belong to the same high-quality image category as the target image and have the same image color distribution characteristics and image contour distribution characteristics; and input each image to be identified into the final form of the approximation program in sequence to obtain the corresponding image feature recognition results;
[0062] The recognition result of each image feature is then input into the target loop program to obtain the final image feature information of each image to be recognized, and to construct a one-to-one mapping relationship between each image to be recognized and its corresponding final image feature information.
[0063] The beneficial effects of the above technical solution are as follows: Through this method, clustered image feature recognition can be performed on several images to be recognized that share similar characteristics. Furthermore, by inputting the image feature recognition result corresponding to each image to be recognized into the target loop program, the final approximate program has already performed preliminary feature recognition analysis on the images to be recognized, enabling the subsequent target loop program to perform secondary optimization of the image feature recognition results in a short time, thereby improving the accuracy and reliability of image feature recognition.
[0064] The beneficial effects of the above technical solution are as follows: This VGG image feature extraction acceleration method based on approximation calculation first selects a target loop program matching the target image from the set of image feature extraction programs based on VGG networks, and performs perforation on the target loop program to obtain an initial approximation program. In this way, the initial approximation program can have the same image recognition function as the target loop program, and the initial approximation program has a higher computational speed than the target loop program. Then, based on the difference between the image feature recognition results of the target image by the target loop program and the initial approximation program, it is determined whether to perform secondary perforation on the initial approximation program. Finally, the image feature recognition result of the final approximation program is input into the target loop program to obtain the final image feature information of each image to be recognized. The above method generates an initial approximation program in a perforated manner for preliminary feature recognition of the image to be recognized, which can reduce the computational workload of the subsequent target loop program and accelerate image feature extraction.
[0065] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method for accelerating VGG image feature extraction based on approximate calculation, characterized in that, It includes the following steps: Step S1: Obtain the target image uploaded by the client terminal, perform preliminary screening on the target image, and determine the image type information of the target image; based on the image type information, select a target loop program that matches the target image from the set of image feature extraction programs based on VGG network; Step S2: Perform perforation on the selected target loop program to obtain an initial approximation program; input the target image into the target loop program and the initial approximation program respectively for analysis and processing, thereby extracting the first image feature information and the second image feature information; Step S3: Based on the difference in image feature recognition accuracy between the first image feature information and the second image feature information, determine whether to perform secondary perforation on the initial approximation procedure; and based on the above determination result, obtain the final form of the approximation procedure; Step S4 involves sequentially inputting several images to be identified that have the same image type information as the target image into the final approximation program, and inputting the image feature recognition result corresponding to each image to be identified into the target loop program, thereby obtaining the final image feature information of each image to be identified, specifically including: Acquire several images that belong to the same high-quality image category as the target image and have the same image color distribution characteristics and image contour distribution characteristics; and input each image to be identified sequentially into the final form of the approximation program to obtain the corresponding image feature recognition result; The image feature recognition result of each image is then input into the target loop program to obtain the final image feature information of each image to be recognized, and to construct a one-to-one mapping relationship between each image to be recognized and its corresponding final image feature information.
2. The accelerated VGG image feature extraction method based on approximate calculation as described in claim 1, characterized in that: In step S1, acquiring the target image uploaded by the client terminal and performing preliminary screening on the target image to determine the image type information specifically includes: The target image uploaded by the client terminal is acquired, and the target image is subjected to image brightness analysis and image sharpness analysis to obtain the image brightness value and image pixel resolution value of the target image. If the image brightness value is greater than or equal to a preset brightness threshold and the image pixel resolution value is greater than or equal to a preset resolution threshold, then the target image is determined to be a high-quality image; otherwise, the target image is determined to be a low-quality image.
3. The accelerated VGG image feature extraction method based on approximate calculation as described in claim 2, characterized in that: In step S1, selecting a target loop program that matches the target image from the set of image feature extraction programs based on the VGG network, according to the image type information, specifically includes: When the target image is a high-quality image, a target loop program that matches the image color distribution features and the image contour distribution features is selected from the set of image feature extraction programs based on VGG networks, according to the image color distribution features and the image contour distribution features of the target image.
4. The accelerated VGG image feature extraction method based on approximate calculation as described in claim 3, characterized in that: In step S2, the selected target loop program is perforated to obtain an initial approximate program, specifically including: The target loop program selected above is subjected to a predetermined number of continuous loop punching operations to obtain an initial approximate program.
5. The accelerated VGG image feature extraction method based on approximate calculation as described in claim 4, characterized in that: In step S2, the target image is input into the target loop program and the initial approximation program for analysis and processing, thereby extracting the first image feature information and the second image feature information. Specifically, this includes: The target image is sequentially pixel-sharpened and then input into the target loop program and the initial approximation program for analysis and processing, thereby extracting the first image feature information and the second image feature information respectively; wherein, the first image feature information and the second image feature information refer to the edge line contour features of people or objects in the target image.
6. The accelerated VGG image feature extraction method based on approximate calculation as described in claim 5, characterized in that: In step S3, determining whether to perform a secondary perforation on the initial approximation procedure based on the difference in image feature recognition accuracy between the first image feature information and the second image feature information specifically includes: Determine the degree of overlap between the first image feature information and the second image feature information in terms of the line contours at the edge lines; If the line contour overlap value is greater than or equal to the prediction degree threshold, it is determined that no secondary perforation is required for the initial approximation procedure; otherwise, it is determined that secondary perforation is required for the initial approximation procedure.
7. The accelerated VGG image feature extraction method based on approximate calculation as described in claim 6, characterized in that: In step S3, the approximate procedure for obtaining the final form based on the above judgment results specifically includes: If it is determined that a second perforation is not required for the initial approximation procedure, then the current initial approximation procedure is used as the final form of the approximation procedure. When it is determined that a second perforation is required for the initial approximation procedure, the initial approximation procedure after at least one round of second perforation is taken as the final form of the approximation procedure, provided that the final line contour overlap value is greater than or equal to the prediction degree threshold.