Multi-model fusion method based on deep learning and its application in wear-resistant plate detection

Abstract

The present invention provides a multi-model fusion method based on deep learning, comprising the following steps: in the model preprocessing stage, the characteristics of the input image data are analyzed, and the interference-causing images and normal images are classified; the interference images are trained using a classification filtering module and a classification model, and the normal images are trained using a segmentation filtering module and a segmentation model, and the corresponding model weight files are obtained and applied to the subsequent application stage; in the model application stage, the classification model is first used to classify the image data, and the images belonging to the filter class are directly output, and the rest are input to the segmentation model for judgment; the data processed by the segmentation model will be screened by an adaptive filtering module, and the final data will be output. The present invention adopts a multi-model fusion approach, which effectively overcomes the shortcomings of a single model algorithm.

Description

Technical Field

[0001] The present invention relates to the field of computer vision in deep learning, and in particular to a multi-model fusion method based on deep learning and its application in wear-resistant plate detection. Background Art

[0002] With the rapid development of deep learning, its related technologies have been gradually applied to various industries, replacing many traditional technologies and achieving superior performance. In the field of industrial defect detection, accurately detecting and distinguishing good from defective products is particularly important. Therefore, research on how to improve model performance in this field is of great significance.

[0003] Industrial defect detection aims to leverage relevant technologies to quickly and accurately locate defects within defective products and ultimately distinguish good from defective ones. Before the rise of deep learning, this approach primarily relied on manual and machine learning methods. These methods were limited by time-consuming labor, high error rates, and the inability of machine learning to meet the demands of higher standards. The application of deep learning has effectively addressed these issues and significantly improved industry efficiency.

[0004] Although deep learning technology has been applied to a certain extent, it still has the following limitations in special scenarios: when a single model encounters some special problems, its problem graph cannot be filtered through post-processing methods, which greatly reduces its accuracy.

[0005] Therefore, it is of great practical significance to develop a multi-model fusion solution that can achieve fast, accurate, low-cost classification of good and defective products with good adaptability.

[0006] Related technologies are currently widely used in inspection tasks across various fields and demonstrate good performance. However, for some specific scenarios, relying solely on a single model cannot achieve the desired results. Therefore, a multi-model fusion solution has been proposed to effectively address this issue. This invention can be applied to wearplate inspection projects and can improve detector performance. Summary of the Invention

[0007] To address the problem that existing single models are unable to solve specific scenarios (and even post-processing methods cannot accurately filter them), this paper aims to effectively solve this problem by proposing a multi-model fusion solution, and to provide some ideas and help for similar problems. This solution will use wear plate detection as an example. First, the image is analyzed based on its style type. Then, the corresponding multi-model fusion solution is selected based on the analysis. Finally, each model performs its own role, accurately detecting while filtering out interference items and finally completing the classification.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The multi-model fusion method based on deep learning includes the following steps:

[0010] In the model preprocessing stage, the characteristics of the input image data are analyzed to classify the images that cause interference from the normal images;

[0011] The interference graph is trained using a classification filtering module and a classification model, and the normal graph is trained using a segmentation filtering module and a segmentation model. After obtaining the corresponding model weight files, they are applied to the subsequent application stage;

[0012] In the model application stage, the classification model is first used to classify the image data, and the images belonging to the filtering class are directly output, and the rest are input into the segmentation model for judgment;

[0013] The data processed by the segmentation model will be filtered by an adaptive filtering module and the final data will be output.

[0014] As a further solution of the present invention, in the model preprocessing stage, the features of the input images are analyzed, and the images that affect the model effect are uniformly divided into interference images, and the rest are regarded as normal images.

[0015] As a further solution of the present invention, the classification filtering module sets a filtering class for each type of interference graph, and transmits the images of the filtering class to the classification model for training; in addition, the normal image passes through the segmentation filtering module, and all its images are transmitted to the segmentation model for training.

[0016] As a further solution of the present invention, in the model application stage, a batch of new data will all be passed to the classification model for verification, and then judgment will be made based on the output results of the classification model. The pictures belonging to the interference class will be directly output, and those that do not belong to their category will be passed to the segmentation model for the next step of operation; after the segmentation model is processed, it will be filtered through an adaptive filtering module, and finally the results with segmentation mask and detection box will be output.

[0017] As a further solution of the present invention, the processing flow of the adaptive filtering module is: obtaining the total number of targets output by the segmentation algorithm, the coordinates of each target output by the segmentation algorithm, the total number of targets threshold, the horizontal coordinate threshold, and finally outputting the filtered results according to the filtering algorithm.

[0018] The present invention also discloses an application of multi-model fusion based on deep learning in wear-resistant plate detection, which includes the following steps:

[0019] In the model preprocessing stage, the features of the input wear-resistant plate images will be analyzed, and the images that affect the model effect will be uniformly divided into interference images, and the rest will be regarded as normal images and trained separately;

[0020] During the model application phase, a batch of new data will be passed to the classification model for verification. Then, a judgment will be made based on the output of the classification model. Pictures belonging to the interference class will be directly output, while those that do not belong to their category will be passed to the segmentation model for the next step. After the segmentation model is processed, it will be filtered through an adaptive filtering module, and finally the result with segmentation mask and detection box will be output.

[0021] As a further solution of the present invention, the pre-set conditions of the adaptive filtering module include: obtaining the total number of targets output by the segmentation algorithm, the coordinates of each target output by the segmentation algorithm, and a threshold value for the total number of targets.

[0022] As a further solution of the present invention, the algorithm of the adaptive filtering module is specifically as follows: determine whether the total number of targets currently detected is greater than the target total number threshold. If it is, enter the subsequent filtering step, otherwise directly output without filtering; calculate the horizontal and vertical coordinates of the center point coordinates of each target based on the upper left and lower right horizontal and vertical coordinates of each target in the target coordinates output by the segmentation algorithm; take the center point coordinates (cx0, cy0) of the center circle of the wear-resistant plate as the reference, compare the horizontal coordinates of the center point coordinates of the remaining points with cx0, and store the center point coordinate values ​​of targets smaller than cx0 in the target data smaller than the center point coordinates of the center circle, and store the center point coordinates of targets greater than cx0 in the target data greater than the center point coordinates of the center circle.

[0023] As a further solution of the present invention, the lengths of target data smaller than the center point coordinates of the center circle and target data larger than the center point coordinates of the center circle are determined respectively, and the lists with lengths greater than 2 are input into the subsequent filtering link.

[0024] As a further solution of the present invention, the center point coordinate value of each target in the current list is calculated to measure the distance with the center point coordinate of the center circle, and the difference values ​​are compared pairwise. The horizontal and vertical coordinates of the target with the larger difference value are stored in the secondary candidate list, and the horizontal and vertical coordinates of the two targets with the smallest difference value are stored in the secondary inspection list; the secondary inspection list is traversed to find the smallest upper left corner vertical coordinate and the largest lower right corner vertical coordinate of the two target points in the current list, representing two points respectively, and the difference values ​​are compared with the smallest vertical coordinate and the largest vertical coordinate of the center circle respectively. As long as the difference between the two points and the vertical coordinate is less than the vertical coordinate threshold, the subsequent second horizontal coordinate threshold judgment is entered, otherwise the data is added to the secondary candidate list and waits for filtering; if the condition that the difference value is less than the vertical coordinate threshold is met, the second horizontal coordinate threshold judgment is performed, and its main steps are: calculate the center point coordinates of the two targets, and then calculate the difference between the horizontal coordinates of the center point coordinates of the two. If it is less than the horizontal coordinate threshold, it is directly used as the final filtered data and output;

[0025] If it is not satisfied, the data in the secondary inspection list will be added to the secondary candidate list, and the above two threshold judgments will be repeated; the final output filtered result includes the horizontal and vertical coordinates of the center circle and the horizontal and vertical coordinates of the four welding points as the final filtered data.

[0026] The present invention has the following beneficial effects:

[0027] 1. The multi-model fusion method is adopted to effectively overcome the shortcomings of the single model algorithm.

[0028] 2. An adaptive filtering module algorithm is proposed to improve the performance of the model.

[0029] 3. Taking wearplate as an example, we provide some ideas for multi-model fusion solutions for similar projects.

[0030] In order to more clearly illustrate the structural features and effects of the present invention, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a wearplate input example in an embodiment of the present invention.

[0032] Figure 2 This is the segmentation result in the embodiment of the present invention, displayed in the form of a picture.

[0033] Figure 3 This is an interference diagram in an embodiment of the present invention, which is a "reverse installation type of iron sheet".

[0034] Figure 4 This is the model preprocessing stage in the embodiment of the present invention.

[0035] Figure 5 This is the model application stage in the embodiment of the present invention.

[0036] Figure 6 、 Figure 7 They are comparison diagrams of the results after filtering by the adaptive filtering module in the embodiments of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described below with reference to the accompanying drawings and related knowledge, and described clearly and completely. Obviously, the described applications are only part of the embodiments of the present invention, rather than all of the embodiments.

[0038] Example 1

[0039] The present invention provides a multi-model fusion method based on deep learning, comprising the following steps:

[0040] In the model preprocessing stage, the characteristics of the input image data are analyzed to classify the images that cause interference from the normal images;

[0041] The interference graph is trained using a classification filtering module and a classification model, and the normal graph is trained using a segmentation filtering module and a segmentation model. After obtaining the corresponding model weight files, they are applied to the subsequent application stage;

[0042] In the model application stage, the classification model is first used to classify the image data, and the images belonging to the filtering class are directly output, and the rest are input into the segmentation model for judgment;

[0043] The data processed by the segmentation model will be filtered by an adaptive filtering module and the final data will be output.

[0044] In this embodiment, during the model preprocessing stage, features of the input images are analyzed, and images that affect the model effect are uniformly classified as interference images, while the rest are classified as normal images.

[0045] In this embodiment, the classification filtering module sets a filtering class for each type of interference graph and transmits the images of the filtering class to the classification model for training; in addition, the normal image passes through the segmentation filtering module and all its images are transmitted to the segmentation model for training.

[0046] In this embodiment, during the model application stage, a batch of new data will all be passed to the classification model for verification, and then a judgment will be made based on the output results of the classification model. The pictures belonging to the interference class will be directly output, and those that do not belong to their category will be passed to the segmentation model for the next step of operation; after the segmentation model is processed, it will be filtered through an adaptive filtering module, and finally the results with segmentation mask and detection box will be output.

[0047] In this embodiment, the processing flow of the adaptive filtering module is as follows: obtaining the total number of targets output by the segmentation algorithm, the coordinates of each target output by the segmentation algorithm, the total number of targets threshold, the horizontal coordinate threshold, and finally outputting the filtered result according to the filtering algorithm.

[0048] In the present invention, a multi-model fusion solution based on deep learning, first, in the model preprocessing stage, according to the characteristics of the input image data, the images that are prone to interference are analyzed and classified from the normal images. Secondly, the interference image is processed using filtering module No. 1 and model No. 1 (here is the classification model), and the normal image is processed using filtering module No. 2 and model No. 2 (here is the segmentation model). Thirdly, in the preprocessing stage, the two types of models are trained separately, and after obtaining the corresponding model weight files, they are applied to the subsequent application stage. From this, in the model application stage, model No. 1 (classification model) is first used to classify the image data, and the images belonging to the filtering class are directly output, and the rest are input to model No. 2 for judgment (segmentation model). Finally, the data processed by model No. 2 will be screened by an adaptive filtering module, and the final data will be output. In summary, the proposed method can solve the limitations of most current single technologies, thereby improving the performance of the model.

[0049] Example 2

[0050] The present invention provides an application of multi-model fusion based on deep learning in wear-resistant plate detection, comprising the following steps:

[0051] In the model preprocessing stage, the features of the input wear-resistant plate images will be analyzed, and the images that affect the model effect will be uniformly divided into interference images, and the rest will be regarded as normal images and trained separately;

[0052] During the model application phase, a batch of new data will be passed to the classification model for verification. Then, a judgment will be made based on the output of the classification model. Pictures belonging to the interference class will be directly output, while those that do not belong to their category will be passed to the segmentation model for the next step. After the segmentation model is processed, it will be filtered through an adaptive filtering module, and finally the result with segmentation mask and detection box will be output.

[0053] In this embodiment, the pre-set conditions of the adaptive filtering module include: obtaining the total number of targets output by the segmentation algorithm, the coordinates of each target output by the segmentation algorithm, and a threshold value for the total number of targets.

[0054] In this embodiment, the algorithm of the adaptive filtering module is specifically as follows: determine whether the total number of targets currently detected is greater than the target total number threshold. If it is, enter the subsequent filtering step, otherwise directly output without filtering; calculate the horizontal and vertical coordinates of the center point coordinates of each target based on the upper left and lower right horizontal and vertical coordinates of each target in the target coordinates output by the segmentation algorithm; take the center point coordinates (cx0, cy0) of the center circle of the wear-resistant plate as the reference, compare the horizontal coordinates of the center point coordinates of the remaining points with cx0, and store the center point coordinate values ​​of targets smaller than cx0 in the target data smaller than the center point coordinates of the center circle, and store the center point coordinates of targets greater than cx0 in the target data greater than the center point coordinates of the center circle.

[0055] In this embodiment, the lengths of target data smaller than the center point coordinates of the center circle and target data larger than the center point coordinates of the center circle are determined respectively, and a list with a length greater than 2 is input into a subsequent filtering step.

[0056] In this embodiment, the center point coordinate value of each target in the current list is calculated to measure the distance with the center point coordinate of the center circle, and the difference values ​​are compared pairwise. The horizontal and vertical coordinates of the target with the larger difference value are stored in the secondary candidate list, and the horizontal and vertical coordinates of the two targets with the smallest difference value are stored in the secondary inspection list; the secondary inspection list is traversed to find the smallest upper left corner vertical coordinate and the largest lower right corner vertical coordinate of the two target points in the current list, representing two points respectively, and the difference values ​​are compared with the smallest vertical coordinate and the largest vertical coordinate of the center circle respectively. As long as the difference between the two points and the vertical coordinate is less than the vertical coordinate threshold, the subsequent second horizontal coordinate threshold judgment is entered, otherwise the data is added to the secondary candidate list and waits for filtering; if the condition that the difference value is less than the vertical coordinate threshold is met, the second horizontal coordinate threshold judgment is performed, and its main steps are: calculate the center point coordinates of the two targets, and then calculate the difference between the horizontal coordinates of the center point coordinates of the two. If it is less than the horizontal coordinate threshold, it is directly used as the final filtered data and output;

[0057] If it is not satisfied, the data in the secondary inspection list will be added to the secondary candidate list, and the above two threshold judgments will be repeated; the final output filtered result includes the horizontal and vertical coordinates of the center circle and the horizontal and vertical coordinates of the four welding points as the final filtered data.

[0058] This solution will take wear plate detection as an example. First, it analyzes the image based on its style type. Then, it selects the corresponding multi-model fusion solution based on the analysis results. Finally, each model performs its own function. While accurately detecting, it will filter out interference items and finally complete the classification.

[0059] Example 3

[0060] The multi-model fusion solution based on deep learning mainly includes the model preprocessing stage and the model application stage. Taking wear plate as an example, as shown in the attached Figure 1 and Figure 2 As shown in the figure, the final output of the wear plate is the center circle target and the four welding points around it. However, there is an interference pattern in the wear plate, as shown in the figure. Figure 3 As shown in the figure, the main characteristic of the "iron patch installed upside down" category is that the center patch is installed upside down. If a single segmentation model algorithm is used, five detection frames will be identified, and the post-processing algorithm will not be able to accurately filter it. Therefore, a multi-model fusion solution is used to solve this problem.

[0061] In the model preprocessing stage, the features of the input image will be analyzed and similar attached Figure 3 Other images that affect the model effect are uniformly classified as interference images, and the rest are regarded as normal images. First, the interference image will pass through the classification filtering module. The main function of this module is to set a filter class for each type of interference image, such as "iron sheet installed upside down", "patch installed crookedly", "no iron sheet" and other types, to provide filtering assistance for the model fusion process in the subsequent application stage. Then all the images with the filter class set will be transferred to model No. 1 (classification model) for training. Similarly, the normal image will pass through the segmentation filtering module, whose main function is the same as the classification filtering module, and all its images will be transferred to model No. 2 (segmentation model) for training.

[0062] During the model application phase, a new batch of data is directly passed to Model 1 (the classification model) for verification. The classification model then makes a judgment based on its output, filtering out all images that fall into the interference category (the filtering category, i.e., the labels assigned by the classification and filtering module). Images that fall into the interference category are directly output, while those that do not fall into this category are passed to Model 2 (the segmentation model) for further processing. After the segmentation model completes processing, it is filtered through an adaptive filtering module (AF) and ultimately outputs a result with a segmentation mask and detection box.

[0063] The main processing flow of the adaptive filtering module (AF) is as follows:

[0064] Prerequisites:

[0065] Get the total number of targets output by the segmentation algorithm (input_count, representing the total number of targets currently detected), the coordinates of each target output by the segmentation algorithm (input_data, each target has upper left and lower right horizontal and vertical coordinate values), the total number of targets threshold (thresh_count=5, the purpose is to use the value of input_count to determine whether adaptive filtering is needed), the vertical coordinate threshold (thresh_y=15, the subsequent indicator for determining whether filtering is needed, used in conjunction with thresh_x), and the horizontal coordinate threshold (thresh_x=20, the subsequent indicator for determining whether filtering is needed, used in conjunction with thresh_y).

[0066] left_list = [ ] # Used to store target data that is smaller than the center point coordinates of the center circle

[0067] right_list = [ ] # Used to store target data that is larger than the center point coordinates of the center circle

[0068] second_verify_list = [ ] # Second verification list

[0069] second_candidate_list = [ ] # Secondary candidate list

[0070] final_result = [ ] # Final output filtered results

[0071] The specific algorithm is:

[0072] Determine whether the value of input_count is greater than thresh_count (here is 5). If True, proceed to the subsequent filtering step, otherwise output directly without filtering.

[0073] The center point coordinates (cx, cy) of each target are calculated based on the upper left and lower right horizontal and vertical coordinates of each target in input_data, representing the horizontal and vertical coordinates of the center point coordinates.

[0074] Using the center point coordinates (cx0, cy0) of the center circle in the wearplate project as a reference, compare the horizontal coordinates of the center point coordinates of the remaining points with cx0. Targets with center point coordinates smaller than cx0 are stored in left_list, while targets with center point coordinates greater than cx0 are stored in right_list. (For example, if five targets are detected excluding the center circle, the total input_count is 6, exceeding the filtering threshold and requiring filtering. Assuming that the target center coordinates pass the filtering, t1, t2, and t3 (the three target center coordinates) are stored in left_list, and t4 and t5 (the two target center coordinates) are stored in right_list).

[0075] Determine the length of left_list and right_list respectively, and make subsequent judgments on the lists whose length is greater than 2.

[0076] Enter the loop and traverse the list currently filtered by step 4 (maybe just one, or maybe both are needed).

[0077] Calculate the distance between the center point coordinates of each target in the current list and the center point coordinates of the center circle, and compare the difference between them. Store the horizontal and vertical coordinates of the target with the larger difference in the secondary candidate list (second_candidate_list), and store the horizontal and vertical coordinates of the two targets with the smallest difference in the secondary verification list (second_verify_list). For example, in the example of step 3, there are two targets t1, t2, and t3 in the left_list. Assume that the center point coordinates of t1 and t2 are close to the center point coordinates of the center circle, while the difference of t3 is larger. Then the upper left and lower right horizontal and vertical coordinates of t1 and t2 (input_data) are stored in the secondary verification list, and t3 is stored in the secondary candidate list.

[0078] Traverse the secondary check list and find the smallest upper left ordinate and the largest lower right ordinate of the two target points in the current list, representing the two points. Compare the difference between these two points and the smallest and largest ordinates of the center circle. If the difference between these two points is less than the ordinate threshold (thresh_y), proceed to the second horizontal threshold judgment. Otherwise, the data is added to the secondary candidate list and awaits filtering.

[0079] If the condition that the difference is less than the vertical coordinate threshold is met in step 7, the second horizontal coordinate threshold (thresh_x) is judged. The main steps are: calculate the center point coordinates of the two targets, and then calculate the horizontal coordinate difference between the two center point coordinates. If it is less than the horizontal coordinate threshold, it is directly used as the final filtered data (final_result) and output.

[0080] If the condition in step 8 is not met, the data in the secondary inspection list is added to the secondary candidate list, and the judgment of the two thresholds in the 7th and 8th times is repeated.

[0081] The final output, final_result, contains the horizontal and vertical coordinates of the center circle and the horizontal and vertical coordinates of the four welding points as the final filtered data.

[0082] The technical principles of the present invention have been described above in conjunction with specific embodiments, which are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention fall within the scope of protection of the present invention. Those skilled in the art will be able to conceive of other specific embodiments of the present invention without inventive effort, and such methods will fall within the scope of protection of the present invention.

Claims

1. A multi-model fusion method based on deep learning, characterized by: The following steps are involved: In the model preprocessing stage, the characteristics of the input image data are analyzed to classify the images that cause interference from the normal images; The interference graph is trained using a classification filtering module and a classification model, and the normal graph is trained using a segmentation filtering module and a segmentation model. After obtaining the corresponding model weight files, they are applied to the subsequent application stage; In the model application stage, the classification model is first used to classify the image data, and the images belonging to the filtering class are directly output, and the rest are input into the segmentation model for judgment; The data processed by the segmentation model will be filtered by an adaptive filtering module and the final data will be output; The algorithm of the adaptive filtering module is specifically as follows: determine whether the total number of targets currently detected is greater than the target total number threshold. If so, enter the subsequent filtering step, otherwise directly output without filtering; calculate the center point coordinate value of each target according to the upper left and lower right horizontal and vertical coordinates of each target in the target coordinates output by the segmentation algorithm, and the horizontal and vertical coordinates of the center point coordinates of the center point coordinates of each target are represented; taking the center point coordinates (cx0, cy0) of the center circle of the wear-resistant plate as the benchmark, compare the horizontal coordinates of the center point coordinates of the remaining points with cx0, store the center point coordinate value of the target less than cx0 in the target data less than the center point coordinate of the center circle, and store the center point coordinates of the target greater than cx0 in the target data greater than the center point coordinate of the center circle; respectively determine the length of the target data less than the center point coordinate of the center circle and the target data greater than the center point coordinate of the center circle, and input the list with a length greater than 2 into the subsequent filtering link; calculate the distance measurement between the center point coordinate value of each target in the current list and the center point coordinate of the center circle, and compare the difference between them two by two, store the horizontal and vertical coordinates of the target with the larger difference value in the secondary candidate list, and store the horizontal and vertical coordinates of the two targets with the smallest difference in the secondary inspection list; Traverse the secondary inspection list and find the smallest upper left vertical coordinate and the largest lower right vertical coordinate of the two target points in the current list, representing two points respectively, and compare the difference with the smallest vertical coordinate and the largest vertical coordinate of the center circle respectively. As long as the difference between the two points and them is less than the vertical coordinate threshold, enter the subsequent second horizontal coordinate threshold judgment, otherwise the data is added to the secondary candidate list and wait for filtering; if the condition that the difference is less than the vertical coordinate threshold is met, the second horizontal coordinate threshold judgment is performed, and its main steps are: calculate the center point coordinates of the two targets, and then calculate the difference between the horizontal coordinates of the center point coordinates of the two. If it is less than the horizontal coordinate threshold, it is directly used as the final filtered data and output; if it is not met, the data in the secondary inspection list is added to the secondary candidate list, and the above two threshold judgments are repeated; the final output filtered result contains the horizontal and vertical coordinates of the center circle and the horizontal and vertical coordinates of the four welding points as the final filtered data.

2. The multi-model fusion method based on deep learning according to claim 1, characterized in that: During the model preprocessing stage, the features of the input images will be analyzed, and the images that affect the model effect will be uniformly divided into interference images, and the rest will be regarded as normal images.

3. The multi-model fusion method based on deep learning according to claim 1, characterized in that: The classification filtering module sets a filtering class for each type of interference image, and transmits the images of the filtering class to the classification model for training; in addition, the normal image passes through the segmentation filtering module, and all its images are transmitted to the segmentation model for training.

4. The multi-model fusion method based on deep learning according to claim 1, characterized in that: During the model application phase, a batch of new data will be passed to the classification model for verification. Then, a judgment will be made based on the output of the classification model. Pictures belonging to the interference class will be directly output, while those that do not belong to their category will be passed to the segmentation model for the next step. After the segmentation model is processed, it will be filtered through an adaptive filtering module, and finally the result with a segmentation mask and detection box will be output.

5. Application of the multi-model fusion method according to claim 1 in wear-resistant plate detection, characterized in that: The following steps are involved: In the model preprocessing stage, the features of the input wear-resistant plate images will be analyzed, and the images that affect the model effect will be uniformly divided into interference images, and the rest will be regarded as normal images and trained separately; During the model application phase, a batch of new data will be passed to the classification model for verification. Then, a judgment will be made based on the output of the classification model. Pictures belonging to the interference class will be directly output, while those that do not belong to their category will be passed to the segmentation model for the next step. After the segmentation model is processed, it will be filtered through an adaptive filtering module, and finally the result with segmentation mask and detection box will be output.

6. The use according to claim 5, characterized in that The pre-set conditions of the adaptive filtering module include: obtaining the total number of targets output by the segmentation algorithm, the coordinates of each target output by the segmentation algorithm, and a threshold value for the total number of targets.

Images