Target detection YOLO algorithm processing system and method based on deep learning

Abstract

The invention discloses a target detection YOLO algorithm processing system and method based on deep learning. The system comprises an input processing module, a feature extraction module, a target detection module, an optimization and training module, a data management and evaluation module and an auxiliary module. The method comprises the following steps: step 1, inputting and processing an image; step 2, feature extraction and fusion; step 3, target detection and correction; and step 4, post-processing and result display. 5, optimizing and training the model; step 6, data management and system maintenance; according to the method, the candidate box is generated through a multi-scale sliding window mechanism, the strategy effectively avoids the problem of key detail loss possibly caused when a high-resolution image is zoomed under a single scale, so that the accuracy of target detection is ensured, and meanwhile, an improved non-maximum suppression algorithm is introduced, so that the accuracy of target detection is improved. According to the algorithm, redundant candidate frames can be inhibited and reduced, and only the optimal candidate frame is reserved as a final detection result.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and in particular to a YOLO algorithm processing system and method for target detection based on deep learning. Background Art

[0002] Object detection is one of the core issues in the field of computer vision. Its task is to find all objects of interest in an image and determine their categories and locations. YOLO is a real-time object detection algorithm. Its core idea is to transform the object detection task into a regression problem and use a single neural network to simultaneously locate and classify the object, thereby achieving real-time and efficient object detection. Object detection technology has a wide range of applications in various fields, including but not limited to autonomous driving, security monitoring, medical image analysis, and industrial manufacturing. Object detection is one of the important research directions in the field of computer vision and has broad application prospects and development potential.

[0003] The existing target detection YOLO algorithm has the following defects: First, when dealing with small target detection, due to the grid-like network structure design of the YOLO algorithm, when a high-resolution image is input, although the inherent details of the image will become clearer as the resolution increases, the YOLO algorithm may lose these key details in the process of scaling the image to a preset size, thereby affecting the detection accuracy of small targets; Second, after the multi-scale prediction mechanism is introduced into the YOLO algorithm, although the detection performance can be improved to a certain extent, it also brings new problems; multiple scales will generate more candidate boxes, and these candidate boxes will often have a large number of overlaps when sharing the same coordinate system, forming redundant candidate boxes; This not only increases the computational complexity of the algorithm, but may also reduce the detection speed and affect the overall performance; Third, when the YOLO algorithm outputs the results, the detection results may be affected by factors such as insufficient training or improper adjustment of model parameters, resulting in inaccurate candidate box positioning or incorrect category judgment; This problem highlights that the algorithm has certain limitations in the accuracy of the detection results, and the current version lacks a secondary correction mechanism for the detection results, making it difficult to further improve the detection accuracy. Summary of the invention

[0004] The purpose of the present invention is to provide a YOLO algorithm processing system and method for target detection based on deep learning to solve the problems raised in the above background technology.

[0005] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a YOLO algorithm processing system for target detection based on deep learning, comprising an input processing module, wherein the input processing module comprises an image input module, a data enhancement module and an image preprocessing module, the input processing module is data-connected with a feature extraction module and an optimization and training module, and the feature extraction module establishes a data connection with the optimization and training module, the feature extraction module is data-connected with a target detection module, and the target detection module establishes a data connection with the optimization and training module, and the target detection module comprises a candidate box generation module, a classification module, a regression module, a non-maximum suppression module, a sliding window module and a detection correction module.

[0006] As a further technical solution of the present invention, the feature extraction module includes a backbone network module, a Conv module, a C3 module, a SPPF module and a feature fusion module.

[0007] As a further technical solution of the present invention, the optimization and training module includes a loss function module, an optimizer module, a learning rate adjustment module and a gradient clipping module. The optimization and training module is data-connected to a data management and evaluation module, and the data management and evaluation module includes a data set management module, a model evaluation module and a performance monitoring module.

[0008] As a further technical solution of the present invention, the target detection module is data-connected with an auxiliary module, and the auxiliary module establishes a data connection with the data management and evaluation module. The auxiliary module includes a log recording module, a model saving and loading module, a hardware configuration management module, a data preprocessing acceleration module, an online learning and updating module, a user interaction interface module, an error handling and recovery module, and a security and privacy protection module.

[0009] As a further technical solution of the present invention, the target detection module is data-connected with a post-processing module, and the post-processing module establishes data connections with the data management and evaluation module and the auxiliary module. The post-processing module includes a border adjustment module, a category probability adjustment module and a result visualization module.

[0010] As a further technical solution of the present invention, the feature fusion module includes an FPN module and a PANet module.

[0011] A method for processing a YOLO algorithm for target detection based on deep learning, including step one, image input and processing; step two, feature extraction and fusion; step three, target detection and correction; step four, post-processing and result display; step five, model optimization and training; step six, data management and system maintenance;

[0012] In the above step 1, the original image to be detected is received by the input processing module, and image enhancement and preprocessing are performed;

[0013] In the above step 2, deep features are extracted from the input image through the feature extraction module, and in the Conv module, the features are further extracted and transformed using the convolution layer, BN layer and activation function, and the feature fusion module is used to fuse the feature maps of different levels together;

[0014] In the above step 3, the image target detection is performed by the target detection module and the secondary correction is performed by the detection correction module;

[0015] In the above step 4, the detection results are post-processed and visualized by the post-processing module;

[0016] In the above step 5, the algorithm model is optimized, trained and adjusted through the optimization and training module;

[0017] In the above step six, system data management and maintenance are performed through the data management and evaluation module and the auxiliary module.

[0018] As a further technical solution of the present invention, in the step three, windows of different sizes are set to slide on the image through the sliding window module, and candidate frames of different scales are generated through the candidate frame generation module. To prevent window redundancy, when the window is about to slide out of the image, the right boundary or the lower boundary of the window is made to coincide with the right side or the lower side of the image. After the sliding is completed, all candidate frames are summarized, the targets in the candidate frames are classified through the classification module, and the positions of the candidate frames are fine-tuned through the regression module to improve the positioning accuracy.

[0019] As a further technical solution of the present invention, in step 3, the overlapping candidate boxes are removed by the non-maximum suppression module, and only the best candidate box is retained as the final detection result. The degree of overlap between the candidate boxes is represented by IOU, and the formula is as follows:

[0020]

[0021] Where α and β represent two overlapping candidate frames. The non-maximum suppression strategy in the original YOLO algorithm first arranges the candidate frames in descending order according to their confidence. When the intersection-over-union (IOU) ratio between a candidate frame with higher confidence and a candidate frame with lower confidence reaches a preset threshold, the candidate frame with higher confidence will suppress the candidate frames with lower confidence in the same category. However, this mechanism is limited to dealing with the overlapping problem of candidate frames in the same category, and no processing is performed on the overlapping candidate frames between different categories. In multi-scale sliding window target detection of high-resolution images, especially for small targets, this limitation may lead to insufficient elimination of redundant candidate frames. In order to optimize this situation, the following improvements are made: when the IOU of two candidate frames is higher than a set threshold, regardless of whether they belong to the same category, the candidate frame with lower score will be directly suppressed, which is helpful to more effectively reduce overlapping and redundant candidate frames in target detection tasks in multi-category, high-resolution and complex scenes, thereby improving the accuracy and efficiency of detection.

[0022] As a further technical solution of the present invention, in step 3, the detection correction module performs secondary correction on the initially detected target result, refines and segments the target area, deeply identifies and determines the significant features of each area, and thus calculates the regional significant value S r In this process, based on the principle that the fuzzy areas in the image and the areas that frequently appear in the image have lower salient features, for the image block P i and all other color blocks P j , compare the two in Lab color space. If P i The color difference with other blocks is large, which means P i is a significant feature, d c P i and P j The color distance, d p P i and P j The spatial distance of the image block P is obtained by i and P j The final distance is as follows:

[0023]

[0024] For any P j If the results of the above formula are all large, it means that P i is significant, select n images with P i Similar blocks calculate image saliency values:

[0025]

[0026] The pixels outside the salient area are weighted by Euclidean distance. The farther the distance, the lower the weight, thereby reducing the saliency value of the background area and further highlighting the salient area in the image.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention generates candidate frames through a multi-scale sliding window mechanism. This strategy effectively avoids the problem of key detail loss that may be caused when scaling high-resolution images at a single scale, thereby ensuring the accuracy of target detection. At the same time, an improved non-maximum suppression algorithm is introduced, which can suppress and reduce redundant candidate frames and only retain the optimal candidate frames as the final detection results. This not only significantly improves the detection speed, but also effectively avoids the negative impact of too many redundant candidate frames on the detection efficiency. At the same time, a secondary correction is performed on the preliminary detection results. By calculating the saliency value of the image, the present invention can highlight those areas with significant features, further optimizing the accuracy and efficiency of the target detection algorithm, which not only enhances the algorithm's ability to recognize key targets, but also improves its robustness in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a system structure diagram of the present invention;

[0029] Figure 2 It is a module architecture diagram of the feature fusion module of the present invention;

[0030] Figure 3 Module architecture diagram of the target detection module of the present invention;

[0031] Figure 4 A module architecture diagram of the optimization and training module of the present invention;

[0032] Figure 5 is a module architecture diagram of the auxiliary module of the present invention;

[0033] Figure 6 is a flow chart of the method of the present invention;

[0034] Figure 7 It is a flow chart of the system operation of the present invention.

[0035] In the figure: 1. Input processing module; 11. Image input module; 12. Data enhancement module; 13. Image preprocessing module; 2. Feature extraction module; 21. Backbone network module; 22. Conv module; 23. C3 module; 24. SPPF module; 25. Feature fusion module; 251. FPN module; 252. PANet module; 3. Target detection module; 31. Candidate box generation module; 32. Classification module; 33. Regression module; 34. Non-maximum suppression module; 35. Sliding window module; 36. Detection correction module; 4. Optimization and training module; 41. Loss function module; 42. Optimization module; 43. Learning rate adjustment module; 44. Gradient clipping module; 5. Data management and evaluation module; 51. Data set management module; 52. Model evaluation module; 53. Performance monitoring module; 6. Auxiliary module; 61. Logging module; 62. Model saving and loading module; 63. Hardware configuration management module; 64. Data preprocessing acceleration module; 65. Online learning and updating module; 66. User interaction interface module; 67. Error handling and recovery module; 68. Security and privacy protection module; 7. Post-processing module; 71. Border adjustment module; 72. Category probability adjustment module; 73. Result visualization module. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] Please see attached Figure 1 -Attached Figure 5An embodiment of the present invention provides: a YOLO algorithm processing system for target detection based on deep learning, comprising an input processing module 1, the input processing module 1 comprising an image input module 11, a data enhancement module 12 and an image preprocessing module 13, the input processing module 1 is data-connected with a feature extraction module 2 and an optimization and training module 4, and the feature extraction module 2 is data-connected with the optimization and training module 4, the feature extraction module 2 is data-connected with a target detection module 3, and the target detection module 3 is data-connected with the optimization and training module 4, the target detection module 3 comprises a candidate box Generation module 31, classification module 32, regression module 33, non-maximum suppression module 34, sliding window module 35 and detection correction module 36; feature extraction module 2 includes backbone network module 21, Conv module 22, C3 module 23, SPPF module 24 and feature fusion module 25. Backbone network module 21 is used to extract deep features from the input image. In Conv module 22, convolution layer, BN layer and activation function are used to further extract and transform features, and C3 module 23 in YOLO algorithm is used to increase network depth and receptive field to improve feature Extraction capability, using SPPF module 24 to achieve spatial invariance and position invariance of input data, and merging feature maps of different levels together through feature fusion module 25; optimization and training module 4 includes loss function module 41, optimizer module 42, learning rate adjustment module 43 and gradient clipping module 44, and the optimization and training module 4 is connected to data management and evaluation module 5, and data management and evaluation module 5 includes data set management module 51, model evaluation module 52 and performance monitoring module 53, loss function module 41 is used to define the loss function of the model, including classification loss, regression loss, etc., optimizer module 42 is used to select and configure optimizer, such as Adam, SGD, etc., for updating model parameters, learning rate adjustment module 43 is used to adjust the learning rate according to the training progress to accelerate training and avoid overfitting, gradient clipping module 44 is used to prevent gradient explosion and ensure the stability of the training process, data set management module 51 is responsible for loading, preprocessing and dividing the data set, model evaluation module 52 is used to evaluate the model using various evaluation indicators, and performance monitoring module 53 is used to monitor the performance indicators of the training process in real time, such as loss value, accuracy, etc.;The target detection module 3 is data-connected with an auxiliary module 6, and the auxiliary module 6 establishes a data connection with the data management and evaluation module 5. The auxiliary module 6 includes a log recording module 61, a model saving and loading module 62, a hardware configuration management module 63, a data preprocessing acceleration module 64, an online learning and updating module 65, a user interaction interface module 66, an error handling and recovery module 67 and a security and privacy protection module 68. The log recording module 61 is used to record key information in the training process, such as loss value, learning rate, etc. The model saving and loading module 62 is used to save and load the trained model. The hardware configuration management module 63 is used to adjust the algorithm parameters according to the hardware configuration to make full use of hardware resources. The data preprocessing acceleration module 64 is used to use GPU and other acceleration hardware to preprocess the input data to improve the processing speed. The online learning and updating module 65 is used to support online learning and updating of the model to adapt to new data and scenarios. The user interaction interface module 66 is used to provide users with a friendly interactive interface to facilitate users to perform model configuration, training and evaluation operations. The error handling and recovery module 67 is used to When an error is encountered during the process, it can automatically process and try to resume training. The security and privacy protection module 68 is used to ensure data security and privacy protection during the processing process. The target detection module 3 is connected to the post-processing module 7, and the post-processing module 7 is connected to the data management and evaluation module 5 and the auxiliary module 6. The post-processing module 7 includes a border adjustment module 71, a category probability adjustment module 72 and a result visualization module 73. The border adjustment module 71 is used to further optimize and adjust the final output border. The category probability adjustment module 72 is used to smooth the probability of the target category to improve the accuracy. The result visualization module 73 is used to display the detection results in the form of an image, including borders, category labels and confidence levels. The feature fusion module 25 includes an FPN module 251 and a PANet module 252. The FPN module 251 fuses feature maps of different levels together through upsampling and downsampling operations to generate a multi-scale feature pyramid. The PANet module 252 adopts an enhanced feature fusion strategy, including top-down and bottom-up feature fusion paths to further enhance the feature fusion effect. ;

[0038] Please see attached Figure 6 -Attached Figure 7 , an embodiment provided by the present invention: a method for processing a YOLO algorithm for target detection based on deep learning, comprising step one, image input and processing; step two, feature extraction and fusion; step three, target detection and correction; step four, post-processing and result display; step five, model optimization and training; step six, data management and system maintenance;

[0039] In the above step 1, the original image to be detected is received by the input processing module 1, and image enhancement and preprocessing are performed;

[0040] In the above step 2, deep features are extracted from the input image by the feature extraction module 2, and the features are further extracted and transformed by using the convolution layer, BN layer and activation function in the Conv module 22, and the feature maps of different levels are fused together by using the feature fusion module 25;

[0041] In the above step three, the target detection module 3 performs image target detection and the detection correction module 36 performs secondary correction. The sliding window module 35 sets windows of different sizes to slide on the image, and the candidate frame generation module 31 generates candidate frames of different scales. To prevent window redundancy, when the window is about to slide out of the image, the right boundary or the lower boundary of the window is made to coincide with the right side or the lower side of the image. After the sliding is completed, all candidate frames are summarized, and the targets in the candidate frames are classified by the classification module 32. The positions of the candidate frames are fine-tuned by the regression module 33 to improve the positioning accuracy. The overlapping candidate frames are removed by the non-maximum suppression module 34, and only the best candidate frame is retained as the final detection result. The degree of overlap between the candidate frames is represented by IOU, and its formula is as follows:

[0042]

[0043] Where α and β represent two overlapping candidate boxes. The non-maximum suppression strategy in the original YOLO algorithm first sorts the candidate boxes in descending order according to their confidence. When the IOU between a candidate box with higher confidence and a candidate box with lower confidence reaches a preset threshold, the candidate box with higher confidence will suppress the candidate boxes with lower confidence in the same category. However, this mechanism is limited to dealing with the overlapping problem of candidate boxes of the same category, and no processing is performed on the overlapping candidate boxes of different categories. This limitation can be used in multi-scale sliding window target detection of high-resolution images, especially for small targets. This may lead to insufficient elimination of redundant candidate frames. In order to optimize this situation, the following improvements are made: when the IOU of two candidate frames is higher than a certain set threshold, regardless of whether they belong to the same category, the candidate frame with a lower score will be directly suppressed, which is helpful to more effectively reduce overlapping and redundant candidate frames in target detection tasks under multi-category, high resolution and complex scenes, thereby improving the accuracy and efficiency of detection. The detection correction module 36 performs secondary correction on the initially detected target results, refines the segmentation of the target area, deeply identifies and determines the significant features of each area, and thus calculates the regional significant value S r In this process, based on the principle that the fuzzy areas in the image and the areas that frequently appear in the image have lower salient features, for the image block P i and all other color blocks P j , compare the two in Lab color space. If P iThe color difference with other blocks is large, which means P i is a significant feature, d c P i and P j The color distance, d p P i and P j The spatial distance of the image block P is obtained by i and P j The final distance is as follows:

[0044]

[0045] For any P j If the results of the above formula are all large, it means that P i is significant, select n images with P i Similar blocks calculate image saliency values:

[0046]

[0047] The pixels outside the salient area are weighted by Euclidean distance. The farther the distance, the lower the weight, thereby reducing the saliency value of the background area and further highlighting the salient area in the image.

[0048] In the above step 4, the detection result is post-processed by the post-processing module 7 and visualized;

[0049] In the above step 5, the algorithm model is optimized, trained and adjusted through the optimization and training module 4;

[0050] In the above step six, the system data management and maintenance are performed through the data management and evaluation module 5 and the auxiliary module 6.

[0051] Based on the above, the advantages of the present invention are: when the present invention is used for target detection, the original image to be detected is first received through the input processing module 1, and image enhancement and preprocessing are performed, specifically, the original image data is received through the image input module 11 and preprocessed into a format suitable for algorithm processing through the image preprocessing module 13, and then various transformations are performed on the input image through the data enhancement module 12 to increase the diversity of the data and improve the generalization ability of the model, and then the deep features are extracted from the input image through the backbone network module 21 in the feature extraction module 2, and in the Conv module 22, the features are further extracted and transformed using the convolution layer, the BN layer and the activation function, and the network depth and receptive field are increased through the C3 module 23 in the YOLO algorithm to improve the feature extraction ability, and the SPPF module 24 is used to realize the spatial invariance and position invariance of the input data, and the feature fusion module 25 is used to fuse the feature maps of different levels together, wherein the FPN module 251 is operated by upsampling and downsampling. The feature maps of different levels are fused together to generate a multi-scale feature pyramid. The PANet module 252 adopts an enhanced feature fusion strategy, including top-down and bottom-up feature fusion paths to further enhance the feature fusion effect. Then, the target detection module 3 performs image target detection and the detection correction module 36 performs secondary correction. The sliding window module 35 sets windows of different sizes to slide on the image, and the candidate frame generation module 31 generates candidate frames of different scales. To prevent window redundancy, when the window is about to slide out of the image, the right boundary or the lower boundary of the window is made to coincide with the right side or the lower side of the image. After the sliding is completed, all candidate frames are summarized, and the targets in the candidate frames are classified by the classification module 32. The position of the candidate frames is fine-tuned by the regression module 33 to improve the positioning accuracy. The overlapping candidate frames are removed by the non-maximum suppression module 34, and only the best candidate frame is retained as the final detection result. The degree of overlap between the candidate frames is represented by IOU, and its formula is as follows:

[0052]

[0053] Where α and β represent two overlapping candidate boxes. The non-maximum suppression strategy in the original YOLO algorithm first sorts the candidate boxes in descending order according to their confidence. When the IOU between a candidate box with higher confidence and a candidate box with lower confidence reaches a preset threshold, the candidate box with higher confidence will suppress the candidate boxes with lower confidence in the same category. However, this mechanism is limited to dealing with the overlapping problem of candidate boxes of the same category, and no processing is performed on the overlapping candidate boxes of different categories. This limitation can be used in multi-scale sliding window target detection of high-resolution images, especially for small targets. This may lead to insufficient elimination of redundant candidate frames. In order to optimize this situation, the following improvements are made: when the IOU of two candidate frames is higher than a certain set threshold, regardless of whether they belong to the same category, the candidate frame with a lower score will be directly suppressed, which is helpful to more effectively reduce overlapping and redundant candidate frames in target detection tasks under multi-category, high resolution and complex scenes, thereby improving the accuracy and efficiency of detection. The detection correction module 36 performs secondary correction on the initially detected target results, refines the segmentation of the target area, deeply identifies and determines the significant features of each area, and thus calculates the regional significant value S r In this process, based on the principle that the fuzzy areas in the image and the areas that frequently appear in the image have lower salient features, for the image block P i and all other color blocks P j , compare the two in Lab color space. If P i The color difference with other blocks is large, which means P i is a significant feature, d c P i and P j The color distance, d p P i and P j The spatial distance of the image block P is obtained by i and P j The final distance is as follows:

[0054]

[0055] For any P j If the results of the above formula are all large, it means that P i is significant, select n images with P i Similar blocks calculate image saliency values:

[0056]

[0057] The pixels outside the salient area are weighted by Euclidean distance. The farther the distance, the lower the weight, thereby reducing the saliency value of the background area and further highlighting the salient area in the image. After the target detection is completed, the detection result is post-processed and visualized by the post-processing module 7. The final output border is further optimized and adjusted by the border adjustment module 71. The category probability adjustment module 72 smoothes the probability of the target category to improve accuracy. The result visualization module 73 displays the detection result in the form of an image, including borders, category labels, and confidence levels. The algorithm model is optimized, trained, and adjusted by the optimization and training module 4. Specifically, the loss function module 41 defines the loss function of the model, including classification loss, regression loss, etc. The optimizer module 42 selects and configures the optimizer, such as Adam, SGD, etc., for updating the model parameters. The learning rate adjustment module 43 adjusts the learning rate according to the training progress to accelerate the training and avoid overfitting. The gradient clipping module 44 is used to prevent gradient explosion and ensure the stability of the training process. Finally, the data management and evaluation module 5 and the auxiliary module 6 are used to manage and maintain the system data. The data set management module 51 is responsible for loading, preprocessing and dividing the data set, the model evaluation module 52 is used to evaluate the model using various evaluation indicators, the performance monitoring module 53 is used to monitor the performance indicators of the training process in real time, such as loss value, accuracy, etc., the logging module 61 is used to record key information in the training process, such as loss value, learning rate, etc., the model saving and loading module 62 is used to save and load the trained model, the hardware configuration management module 63 is used to adjust the algorithm parameters according to the hardware configuration to make full use of hardware resources, the data preprocessing acceleration module 64 is used to use GPU and other acceleration hardware to preprocess the input data to improve the processing speed, the online learning and updating module 65 is used to support the online learning and updating of the model to adapt to new data and scenarios, the user interaction interface module 66 is used to provide users with a friendly interactive interface to facilitate users to perform operations such as model configuration, training and evaluation, the error handling and recovery module 67 is used to automatically handle and try to resume training when errors are encountered during training, and the security and privacy protection module 68 is used to ensure data security and privacy protection during the processing.

[0058] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A YOLO algorithm processing system for target detection based on deep learning, comprising an input processing module (1), characterized in that: The input processing module (1) comprises an image input module (11), a data enhancement module (12) and an image preprocessing module (13); the input processing module (1) is data-connected to a feature extraction module (2) and an optimization and training module (4), and the feature extraction module (2) establishes a data connection with the optimization and training module (4); the feature extraction module (2) is data-connected to a target detection module (3), and the target detection module (3) establishes a data connection with the optimization and training module (4); the target detection module (3) comprises a candidate frame generation module (31), a classification module (32), a regression module (33), a non-maximum suppression module (34), a sliding window module (35) and a detection correction module (36).

2. According to a deep learning-based target detection YOLO algorithm processing system according to claim 1, it is characterized in that: The feature extraction module (2) comprises a backbone network module (21), a Conv module (22), a C3 module (23), an SPPF module (24) and a feature fusion module (25).

3. The target detection YOLO algorithm processing system based on deep learning according to claim 1, characterized in that: The optimization and training module (4) comprises a loss function module (41), an optimizer module (42), a learning rate adjustment module (43) and a gradient clipping module (44); the optimization and training module (4) is data-connected to a data management and evaluation module (5); the data management and evaluation module (5) comprises a data set management module (51), a model evaluation module (52) and a performance monitoring module (53).

4. The target detection YOLO algorithm processing system based on deep learning according to claim 1, characterized in that: The target detection module (3) is data-connected with an auxiliary module (6), and the auxiliary module (6) establishes a data connection with the data management and evaluation module (5), and the auxiliary module (6) includes a log recording module (61), a model saving and loading module (62), a hardware configuration management module (63), a data preprocessing acceleration module (64), an online learning and updating module (65), a user interaction interface module (66), an error handling and recovery module (67) and a security and privacy protection module (68).

5. The target detection YOLO algorithm processing system based on deep learning according to claim 4, characterized in that: The target detection module (3) is data-connected to a post-processing module (7), and the post-processing module (7) is data-connected to both the data management and evaluation module (5) and the auxiliary module (6). The post-processing module (7) includes a border adjustment module (71), a category probability adjustment module (72), and a result visualization module (73).

6. The target detection YOLO algorithm processing system based on deep learning according to claim 2, characterized in that: The feature fusion module (25) includes an FPN module (251) and a PANet module (252).

7. A method for processing a YOLO algorithm for target detection based on deep learning, comprising step 1, image input and processing; step 2, feature extraction and fusion; step 3, target detection and correction; step 4, post-processing and result display; step 5, model optimization and training; step 6, data management and system maintenance; characterized in that: In the above step 1, the original image to be detected is received by the input processing module (1), and image enhancement and preprocessing are performed; In the above step 2, deep features are extracted from the input image by the feature extraction module (2), the features are further extracted and transformed by using the convolution layer, the BN layer and the activation function in the Conv module (22), and the feature maps of different levels are fused together by using the feature fusion module (25); In the above step 3, the target detection module (3) performs image target detection and the detection correction module (36) performs secondary correction; In the above step 4, the detection result is post-processed and visualized by the post-processing module (7); In the above step 5, the algorithm model is optimized, trained and adjusted through the optimization and training module (4); In the above step six, the system data management and maintenance are performed through the data management and evaluation module (5) and the auxiliary module (6).

8. The method for processing a target detection YOLO algorithm based on deep learning according to claim 7, characterized in that: In the step three, windows of different sizes are set to slide on the image through the sliding window module (35), and candidate frames of different scales are generated through the candidate frame generation module (31). To prevent window redundancy, when the window is about to slide out of the image, the right boundary or the lower boundary of the window is made to coincide with the right side or the lower side of the image. After the sliding is completed, all candidate frames are summarized, and the objects in the candidate frames are classified through the classification module (32), and the positions of the candidate frames are fine-tuned through the regression module (33) to improve the positioning accuracy.

9. The method for processing a target detection YOLO algorithm based on deep learning according to claim 8, characterized in that: In step 3, the overlapping candidate boxes are removed by the non-maximum suppression module (34), and only the best candidate box is retained as the final detection result. The degree of overlap between the candidate boxes is represented by IOU, and its formula is as follows: Where α and β represent two overlapping candidate frames. The non-maximum suppression strategy in the original YOLO algorithm first arranges the candidate frames in descending order according to their confidence. When the intersection-over-union (IOU) ratio between a candidate frame with higher confidence and a candidate frame with lower confidence reaches a preset threshold, the candidate frame with higher confidence will suppress the candidate frames with lower confidence in the same category. However, this mechanism is limited to dealing with the overlapping problem of candidate frames in the same category, and no processing is performed on the overlapping candidate frames between different categories. In multi-scale sliding window target detection of high-resolution images, especially for small targets, this limitation may lead to insufficient elimination of redundant candidate frames. In order to optimize this situation, the following improvements are made: when the IOU of two candidate frames is higher than a set threshold, regardless of whether they belong to the same category, the candidate frame with lower score will be directly suppressed, which is helpful to more effectively reduce overlapping and redundant candidate frames in target detection tasks in multi-category, high-resolution and complex scenes, thereby improving the accuracy and efficiency of detection.

10. The method for processing a target detection YOLO algorithm based on deep learning according to claim 8, characterized in that: In step 3, the detection correction module (36) performs secondary correction on the initially detected target result, refines and segments the target area, deeply identifies and determines the significant features of each area, and thus calculates the regional significant value S r In this process, based on the principle that the fuzzy areas in the image and the areas that frequently appear in the image have lower salient features, for the image block P i and all other color blocks P j , compare the two in Lab color space. If P i The color difference with other blocks is large, which means P i is a significant feature, d c P i and P j The color distance, d p P i and P j The spatial distance of the image block P is obtained by i and P j The final distance is as follows: For any P j If the results of the above formula are all large, it means that P i is significant, select n images with P i Similar blocks calculate image saliency values: The pixels outside the salient area are weighted by Euclidean distance. The farther the distance, the lower the weight, thereby reducing the saliency value of the background area and further highlighting the salient area in the image.

Images