Method for removing salt mist interference in visual abnormal state detection of circuit board

Abstract

A method for removing salt fog interference in circuit board visual abnormal state detection comprises the following steps: firstly, selecting a circuit board salt fog picture, and then, carrying out convolution on the circuit board salt fog picture by a front convolution layer module to generate a feature map; the encoder module group is used for carrying out convolution on the feature map to obtain a feature tensor without salt fog features, the convolution module group is used for carrying out convolution on the feature map to generate a new feature map, and the feature map and the feature tensor are fused to obtain the new feature map; the convolution module group is used for carrying out convolution on the feature map to generate a new feature map, the feature map and the feature tensor are fused to generate the new feature map, and the output module is used for carrying out convolution and restoration on the feature map and outputting the feature map as a circuit board picture; salt mist interference is removed through the encoder module group and the convolution module group, and details of the circuit board are restored, so that the subsequent detection accuracy is improved. Therefore, the design can remove the interference of salt mist, and facilitates the detection of the visual abnormal state of the circuit board.

Description

Technical Field

[0001] The invention relates to a method for removing salt fog interference, belongs to the field of circuit board visual state abnormality detection, and in particular to a method for removing salt fog interference in circuit board visual abnormality state detection. Background Art

[0002] In order to test the corrosion resistance of the circuit board, the circuit board will be subjected to a salt spray test, and then the circuit board will be subjected to a visual abnormality test to determine whether the corrosion resistance of the circuit board meets the standards.

[0003] The Chinese patent application number is 202410599267.0, and the application date is May 15, 2024. It discloses a quality system for quickly finding abnormalities, which involves the quality control field of circuit boards. While standardizing on-site management, the real-time link between the SPC unit and the on-site analysis results is used to quickly lock the abnormal location and range, realize real-time monitoring, real-time warning, and rapid processing. The on-site information is input to the SPC unit in real time through the server. Through the KPI unit, the data can be quickly analyzed and processed, effectively realizing real-time monitoring, real-time warning, and rapid processing. Although the detection data of this design includes salt spray data, it still has the following defects: This design does not remove the interference of salt spray during the salt spray test of the circuit board. The salt spray will seriously obscure the real information of the circuit board, making the structure and details of the circuit board unrecognizable, thereby reducing the subsequent detection accuracy.

[0004] The information disclosed in this background technology section is only intended to increase the understanding of the overall background of the application, and should not be regarded as acknowledging or suggesting in any form that the information constitutes the prior art already known to ordinary technicians in this field. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects and problems of the prior art that salt mist interference cannot be removed, and to provide a method for removing salt mist interference in circuit board visual abnormal state detection, which can remove salt mist interference.

[0006] To achieve the above objectives, the technical solution of the present invention is: A method for removing salt spray interference in circuit board visual abnormal state detection, the method comprising the following steps: Step 1: First select a circuit board salt spray picture taken in the salt spray test; Step 2: The front convolutional layer module first convolves the circuit board salt spray image to extract low-dimensional features in the circuit board salt spray image, and then the front convolutional module generates a feature map A based on the aforementioned low-dimensional features; Step 3: First, the encoder module in the encoder module group convolves the feature map A to extract the high-dimensional features of the feature map A with the salt spray features removed, and then the encoder module generates a feature tensor based on the high-dimensional features. Step 4: First, the middle convolution layer module in the convolution module group convolves the feature map A to extract the low-dimensional features in the feature map A, and then the middle convolution layer module generates the feature map A1 according to the aforementioned low-dimensional features, and then the feature map A1 is fused with the feature tensor to obtain the feature map A2, and then the post-convolution layer module in the convolution module group convolves the feature map A2 to extract and enhance the low-dimensional features in the feature map A2, and then the post-convolution layer module generates the feature map A3 according to the aforementioned low-dimensional features; Step 5: first fuse feature map A3 with the feature tensor to enhance the high-dimensional features in feature map A3, and then generate feature map A4. Then the output module convolves feature map A4 to extract the features of feature map A4. Then the output module generates feature map A5 according to the aforementioned features. Then the output module restores feature map A5 to a circuit board image. Then the output module outputs the circuit board image. Thus, salt spray interference is removed and this method ends.

[0007] In the fourth step, the convolution of the feature map A by the middle convolution layer module in the convolution module group is that the multiple middle convolution layer modules in the convolution module group sequentially convolve the feature map A.

[0008] In the fourth step, the multiple middle convolution layer modules in the convolution module group sequentially convolve the feature map A by first convolving the feature map A by the first middle convolution layer module, then convolving the feature map A by the second middle convolution layer module, and then convolving the feature map A by the third middle convolution layer module.

[0009] In the fourth step, the first convolutional layer module first performs convolution on the feature map A, the second convolutional layer module then performs convolution on the feature map A, and then the third convolutional layer module performs convolution on the feature map A. The first convolutional layer module first performs convolution on the feature map A, the second convolutional layer module with more channels then performs convolution on the feature map A, and then the third convolutional layer module with more channels performs convolution on the feature map A.

[0010] The first convolutional layer module first performs convolution on the feature map A, and then the second convolutional layer module with more channels performs convolution on the feature map A, and then the third convolutional layer module with more channels performs convolution on the feature map A. The first convolutional layer module with sixty-four channels first performs convolution on the feature map A, and then the second convolutional layer module with one hundred and twenty-eight channels performs convolution on the feature map A, and then the third convolutional layer module with two hundred and fifty-six channels performs convolution on the feature map A.

[0011] In the fourth step, the convolution of the feature map A2 by the post-convolution layer module in the convolution module group is the convolution of the feature map A2 by the post-convolution layer module with a channel number of 512 in the convolution module group.

[0012] In the third step, the convolution of the feature map A by the encoder module in the encoder module group is that the multiple encoder modules in the encoder module group sequentially convolve the feature map A.

[0013] In the third step, the multiple encoder modules in the encoder module group sequentially convolve the feature map A, firstly the first encoder module in the encoder module group convolves the feature map A, then the second encoder module in the encoder module group convolves the feature map A, and then the third encoder module in the encoder module group convolves the feature map A.

[0014] In the fifth step, the output module performs convolution on the feature map A4 to extract the features of the feature map A4, and then the output module generates the feature map A5 according to the aforementioned features. The output convolution module in the output module first performs convolution on the feature map A4 to extract the features of the feature map A4, and then the output convolution module generates the feature map A5. In the fifth step, the output module restores the feature map A5 to a circuit board image, and then the output module outputs the circuit board image as a fully connected layer module in the output module. The feature map A5 is restored to a circuit board image, and then the fully connected layer module outputs the circuit board image.

[0015] In the fifth step, the feature map A5 is generated by the output convolution module, and the feature map A5 is first generated by the output convolution module, and then the pooling layer module aggregates the feature map A5 to reduce the size and calculation amount of the feature map A5.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. In a method for removing salt spray interference in circuit board visual abnormal state detection of the present invention, the method comprises the following steps: the first step: first select a circuit board salt spray picture taken in a salt spray test; the second step: first convolve the circuit board salt spray picture by the front convolution layer module to generate a feature map A; the third step: first convolve the feature map A by the encoder module in the encoder module group to generate a feature tensor; the fourth step: first convolve the feature map A by the middle convolution module in the convolution module group to generate a feature map A1, then fuse the feature map A1 with the feature tensor to obtain a feature map A2, and then convolve the feature map A2 by the back convolution module to generate a feature map A3; the fifth step: first fuse the feature map A3 with the feature tensor to generate a feature map A4, then convolve the feature map A4 by the output module to obtain a feature map A5, and then restore the feature map A5 to a circuit board picture by the output module and output it, thereby removing the salt spray interference, and then the circuit board picture is detected by the visual abnormal state detection system to judge the quality of the circuit board. The advantages of the present invention also include: First point: extract high-dimensional features from feature map A through the encoder module group to obtain a feature tensor with salt spray features removed, and then fuse the feature tensor with the feature map containing low-dimensional features, so that the real information of the circuit board image can be reflected. At this time, the structure and detail information on the circuit board image can be clearly identified, so the accuracy of subsequent circuit board visual abnormal state detection is improved; Second point: The post-convolution module convolves the feature map A2 to generate the feature map A3. At this time, the low-dimensional features are enhanced; the feature map A3 is fused with the feature tensor to generate the feature map A4. After convolution, the feature map A5 is obtained. At this time, the high-dimensional features are enhanced; therefore, both the high-dimensional features and the low-dimensional features are enhanced to further improve the clarity of the image; Therefore, the present invention can remove the interference of salt mist and is beneficial to the visual abnormal state detection of the circuit board.

[0017] 2. In a method for removing salt fog interference in circuit board visual abnormal state detection of the present invention, in the fourth step, the convolution layer module in the convolution module group performs convolution on the feature map A by first performing convolution on the feature map A by a first convolution layer module with a channel number of sixty-four, then performing convolution on the feature map A by a second convolution layer module with a channel number of one hundred and twenty-eight, and then performing convolution on the feature map A by a third convolution layer module with a channel number of two hundred and fifty-six, performing convolution on the feature map A multiple times, and the number of channels during convolution gradually increases to obtain more low-dimensional features, thereby enriching the details when restoring the circuit board image. Therefore, the present invention is rich in details.

[0018] 3. In a method for removing salt fog interference in circuit board visual abnormal state detection of the present invention, in the third step, the encoder module in the encoder module group convolves the feature map A by first convolving the feature map A by the first encoder module in the encoder module group, then convolving the feature map A by the second encoder module in the encoder module group, and then convolving the feature map A by the third encoder module in the encoder module group, and multiple encoder modules extract high-dimensional features from the feature map A in turn, so more high-dimensional features can be obtained, and the defogging performance can be improved. Therefore, the defogging performance of the present invention is better.

[0019] 4. In a method for removing salt fog interference in circuit board visual abnormal state detection of the present invention, the output module convolves the feature map A4 to generate the feature map A5, and then the output module restores the feature map A5 and outputs it as a circuit board picture. The output convolution module convolves the feature map A4 to generate the feature map A5, and then the pooling layer module aggregates the feature map A4, and then the fully connected layer module restores the feature map A5 and outputs it as a circuit board picture. Through the pooling operation of the pooling layer module, the feature map A5 is aggregated, thereby reducing the size of the feature map A5 and reducing the amount of calculation, thereby accelerating the processing speed. Therefore, the processing speed of the present invention is faster. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of the present invention.

[0021] Figure 2 yes Figure 1 Schematic diagram of the structure.

[0022] Figure 3 This is a flow chart of Example 2.

[0023] Figure 4 This is a flow chart of Example 3.

[0024] Figure 5 This is a flow chart of Example 4.

[0025] Figure 6 yes Figure 5 Schematic diagram of the structure.

[0026] Figure 7 yes Figure 4 Schematic diagram of the structure.

[0027] Figure 8 yes Figure 3 Schematic diagram of the structure.

[0028] Fig. 9 It is a schematic diagram of the application results of the present invention.

[0029] Fig.10It is a schematic diagram of the circuit board salt spray picture.

[0030] Fig.11 It is a schematic diagram of the circuit board picture. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0032] See also Figure 1 — Fig.11 , a method for removing salt spray interference in circuit board visual abnormal state detection, the method comprising the following steps: Step 1: First select a circuit board salt spray picture taken in the salt spray test; Step 2: The front convolutional layer module first convolves the circuit board salt spray image to extract low-dimensional features in the circuit board salt spray image, and then the front convolutional module generates a feature map A based on the aforementioned low-dimensional features; Step 3: First, the encoder module in the encoder module group convolves the feature map A to extract the high-dimensional features of the feature map A with the salt spray features removed, and then the encoder module generates a feature tensor based on the high-dimensional features. Step 4: First, the middle convolution layer module in the convolution module group convolves the feature map A to extract the low-dimensional features in the feature map A, and then the middle convolution layer module generates the feature map A1 according to the aforementioned low-dimensional features, and then the feature map A1 is fused with the feature tensor to obtain the feature map A2, and then the post-convolution layer module in the convolution module group convolves the feature map A2 to extract and enhance the low-dimensional features in the feature map A2, and then the post-convolution layer module generates the feature map A3 according to the aforementioned low-dimensional features; Step 5: first fuse feature map A3 with the feature tensor to enhance the high-dimensional features in feature map A3, and then generate feature map A4. Then the output module convolves feature map A4 to extract the features of feature map A4. Then the output module generates feature map A5 according to the aforementioned features. Then the output module restores feature map A5 to a circuit board image. Then the output module outputs the circuit board image. Thus, salt spray interference is removed and this method ends.

[0033] In the fourth step, the convolution of the feature map A by the middle convolution layer module in the convolution module group is that the multiple middle convolution layer modules in the convolution module group sequentially convolve the feature map A.

[0034] In the fourth step, the multiple middle convolution layer modules in the convolution module group sequentially convolve the feature map A by first convolving the feature map A by the first middle convolution layer module, then convolving the feature map A by the second middle convolution layer module, and then convolving the feature map A by the third middle convolution layer module.

[0035] In the fourth step, the first convolutional layer module first performs convolution on the feature map A, the second convolutional layer module then performs convolution on the feature map A, and then the third convolutional layer module performs convolution on the feature map A. The first convolutional layer module first performs convolution on the feature map A, the second convolutional layer module with more channels then performs convolution on the feature map A, and then the third convolutional layer module with more channels performs convolution on the feature map A.

[0036] The first convolutional layer module first performs convolution on the feature map A, and then the second convolutional layer module with more channels performs convolution on the feature map A, and then the third convolutional layer module with more channels performs convolution on the feature map A. The first convolutional layer module with sixty-four channels first performs convolution on the feature map A, and then the second convolutional layer module with one hundred and twenty-eight channels performs convolution on the feature map A, and then the third convolutional layer module with two hundred and fifty-six channels performs convolution on the feature map A.

[0037] In the fourth step, the convolution of the feature map A2 by the post-convolution layer module in the convolution module group is the convolution of the feature map A2 by the post-convolution layer module with a channel number of 512 in the convolution module group.

[0038] In the third step, the convolution of the feature map A by the encoder module in the encoder module group is that the multiple encoder modules in the encoder module group sequentially convolve the feature map A.

[0039] In the third step, the multiple encoder modules in the encoder module group sequentially convolve the feature map A, firstly the first encoder module in the encoder module group convolves the feature map A, then the second encoder module in the encoder module group convolves the feature map A, and then the third encoder module in the encoder module group convolves the feature map A.

[0040] In the fifth step, the output module performs convolution on the feature map A4 to extract the features of the feature map A4, and then the output module generates the feature map A5 according to the aforementioned features. The output convolution module in the output module first performs convolution on the feature map A4 to extract the features of the feature map A4, and then the output convolution module generates the feature map A5. In the fifth step, the output module restores the feature map A5 to a circuit board image, and then the output module outputs the circuit board image as a fully connected layer module in the output module. The feature map A5 is restored to a circuit board image, and then the fully connected layer module outputs the circuit board image.

[0041] In the fifth step, the feature map A5 is generated by the output convolution module, and the feature map A5 is first generated by the output convolution module, and then the pooling layer module aggregates the feature map A5 to reduce the size and calculation amount of the feature map A5.

[0042] The supplementary description of the present invention is as follows: The effect of the present invention is improved as follows: Fig.10 In the salt spray test, there is a lot of salt spray above the circuit board, which interferes with the photographing of the circuit board. In addition, affected by different ambient atmospheric light (such as evenly scattered sunlight or point light sources of street lamps at night) and different scene depths (such as fog near and fog far away), the circuit board salt spray picture will also have fog with different distributions; please refer to Fig.11 , which is a schematic diagram of a circuit board image processed by the present invention. In this figure, the interference of salt spray is removed, and the details of the circuit board are clearly visible, which is conducive to the subsequent visual abnormality detection of the circuit board; please refer to Fig. 9 The peak signal-to-noise ratio of the present invention is 38.13, and that of the prior art is 31.27. Compared with the prior art, the peak signal-to-noise ratio of the present invention is improved to a certain extent, that is, the available information in the circuit board image is improved; the structural similarity of the present invention is 0.959, and the structural similarity in the prior art is 0.933. Compared with the prior art, the structural similarity of the present invention is improved to a certain extent, that is, the structural information of the circuit board in the circuit board image is improved; the fault monitoring accuracy of the present invention is 94.38%, and that of the prior art is 87.22%. Compared with the prior art, the fault detection accuracy of the present invention is improved to a certain extent, that is, the subsequent detection accuracy of visual abnormalities of the circuit board is improved.

[0043] Embodiment 1: See also Figure 1 — Fig.11 , a method for removing salt spray interference in circuit board visual abnormal state detection, the method comprising the following steps: Step 1: First select a circuit board salt spray picture taken in the salt spray test; Step 2: The front convolutional layer module first convolves the circuit board salt spray image to extract low-dimensional features in the circuit board salt spray image, and then the front convolutional module generates a feature map A based on the aforementioned low-dimensional features; Step 3: First, the encoder module in the encoder module group convolves the feature map A to extract the high-dimensional features of the feature map A with the salt spray features removed, and then the encoder module generates a feature tensor based on the high-dimensional features. Step 4: First, the middle convolution layer module in the convolution module group convolves the feature map A to extract the low-dimensional features in the feature map A, and then the middle convolution layer module generates the feature map A1 according to the aforementioned low-dimensional features, and then the feature map A1 is fused with the feature tensor to obtain the feature map A2, and then the post-convolution layer module in the convolution module group convolves the feature map A2 to extract and enhance the low-dimensional features in the feature map A2, and then the post-convolution layer module generates the feature map A3 according to the aforementioned low-dimensional features; Step 5: first fuse feature map A3 with the feature tensor to enhance the high-dimensional features in feature map A3, and then generate feature map A4. Then the output module convolves feature map A4 to extract the features of feature map A4. Then the output module generates feature map A5 according to the aforementioned features. Then the output module restores feature map A5 to a circuit board image. Then the output module outputs the circuit board image. Thus, salt spray interference is removed and this method ends.

[0044] Embodiment 2: The basic content is the same as that of Example 1, except that: See also Figure 1 — Figure 8 , in the fourth step, the convolution of feature map A by the middle convolution layer module in the convolution module group is that the multiple middle convolution layer modules in the convolution module group sequentially convolve feature map A. In the fourth step, the convolution of feature map A by the multiple middle convolution layer modules in the convolution module group sequentially is that the first middle convolution layer module first convolves feature map A, then the second middle convolution layer module convolves feature map A, and then the third middle convolution layer module convolve feature map A. In the fourth step, the convolution of feature map A by the first middle convolution layer module first convolutes feature map A, then the second middle convolution layer module convolutes feature map A, and then the third middle convolution layer module convolutes feature map A. is that the first middle convolution layer module first convolutes feature map A, then the second middle convolution layer module with more channels convolutes feature map A, and then the third middle convolution layer module with more channels convolutes feature map A. The first convolutional layer module first performs convolution on the feature map A, the second convolutional layer module with more channels performs convolution on the feature map A, and the third convolutional layer module with more channels performs convolution on the feature map A, which means that the first convolutional layer module with sixty-four channels performs convolution on the feature map A, the second convolutional layer module with one hundred and twenty-eight channels performs convolution on the feature map A, and the third convolutional layer module with two hundred and fifty-six channels performs convolution on the feature map A. In the fourth step, the post-convolutional layer module in the convolutional module group performs convolution on the feature map A2, which means that the post-convolutional layer module with five hundred and twelve channels in the convolutional module group performs convolution on the feature map A2.

[0045] When applied, in the fourth step, the feature map A is first convolved by the first middle convolution layer module with a channel number of sixty-four, and then the feature map A is convolved by the second middle convolution layer module with a channel number of one hundred and twenty-eight, and then the feature map A is convolved by the third middle convolution layer module with a channel number of two hundred and fifty-six. At this time, the collected low-dimensional features gradually increase, which is conducive to the subsequent restoration of the details of the circuit board image. The three middle convolution layer modules are the optimal choice after the experiment; then the feature map A1 is generated, and then the feature map A1 is fused with the feature tensor to obtain the feature map A2, and then the feature map A2 is convolved by the post-convolution layer module with a channel number of five hundred and twelve to extract and enhance the low-dimensional features in the feature map A2, and then the post-convolution layer module generates the feature map A3 according to the aforementioned low-dimensional features.

[0046] Embodiment 3: The basic content is the same as that of Example 1, except that: See also Figure 1 — Figure 7 In the third step, the convolution of feature map A by the encoder module in the encoder module group is that the multiple encoder modules in the encoder module group sequentially convolve feature map A. In the third step, the convolution of feature map A by the multiple encoder modules in the encoder module group sequentially is that the first encoder module in the encoder module group first convolves feature map A, then the second encoder module in the encoder module group convolves feature map A, and then the third encoder module in the encoder module group convolves feature map A.

[0047] When applied, in the third step, the first encoder module first convolves the feature map A, then the second encoder module convolves the feature map A, and then the third encoder module convolves the feature map A, and the high-dimensional features of the feature map A with the salt spray features removed are extracted multiple times, and then the feature tensor is generated based on the aforementioned high-dimensional features; the three encoder modules are the optimal choice after the experiment.

[0048] Embodiment 4: The basic content is the same as that of Example 1, except that: See also Figure 1 — Figure 6In the fifth step, the output module performs convolution on the feature map A4 to extract the features of the feature map A4, and then the output module generates the feature map A5 according to the aforementioned features. The output convolution module in the output module first performs convolution on the feature map A4 to extract the features of the feature map A4, and then the output convolution module generates the feature map A5; in the fifth step, the output module restores the feature map A5 to a circuit board image, and then the output module outputs the circuit board image. The fully connected layer module in the output module restores the feature map A5 to a circuit board image, and then the fully connected layer module outputs the circuit board image. In the fifth step, the output convolution module generates the feature map A5. The output convolution module first generates the feature map A5, and then the pooling layer module aggregates the feature map A5 to reduce the size and computational complexity of the feature map A5.

[0049] When applied, in the fifth step, the output convolution module first convolves the feature map A4 to extract the features of the feature map A4, and then the output convolution module generates the feature map A5, and then the pooling layer module aggregates the feature map A5 to reduce the size and calculation amount of the feature map A5, and then the fully connected layer module restores the feature map A5 to a circuit board image, and then the fully connected layer module outputs the circuit board image.

[0050] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.

Claims

1. A method for removing salt spray interference in circuit board visual abnormal state detection, characterized in that: The method comprises the following steps: Step 1: First select a circuit board salt spray picture taken in the salt spray test; Step 2: The front convolutional layer module first convolves the circuit board salt spray image to extract low-dimensional features in the circuit board salt spray image, and then the front convolutional module generates a feature map A based on the aforementioned low-dimensional features; Step 3: First, the encoder module in the encoder module group convolves the feature map A to extract the high-dimensional features of the feature map A with the salt spray features removed, and then the encoder module generates a feature tensor based on the high-dimensional features. Step 4: First, the middle convolution layer module in the convolution module group convolves the feature map A to extract the low-dimensional features in the feature map A, and then the middle convolution layer module generates the feature map A1 according to the aforementioned low-dimensional features, and then the feature map A1 is fused with the feature tensor to obtain the feature map A2, and then the post-convolution layer module in the convolution module group convolves the feature map A2 to extract and enhance the low-dimensional features in the feature map A2, and then the post-convolution layer module generates the feature map A3 according to the aforementioned low-dimensional features; Step 5: first fuse feature map A3 with the feature tensor to enhance the high-dimensional features in feature map A3, and then generate feature map A4. Then the output module convolves feature map A4 to extract the features of feature map A4. Then the output module generates feature map A5 according to the aforementioned features. Then the output module restores feature map A5 to a circuit board image. Then the output module outputs the circuit board image. Thus, salt spray interference is removed and this method ends.

2. A method for removing salt spray interference in circuit board visual abnormal state detection according to claim 1, characterized in that: In the fourth step, the convolution of the feature map A by the middle convolution layer module in the convolution module group is that the multiple middle convolution layer modules in the convolution module group sequentially convolve the feature map A.

3. A method for removing salt spray interference in circuit board visual abnormal state detection according to claim 2, characterized in that: In the fourth step, the multiple middle convolution layer modules in the convolution module group sequentially convolve the feature map A by first convolving the feature map A by the first middle convolution layer module, then convolving the feature map A by the second middle convolution layer module, and then convolving the feature map A by the third middle convolution layer module.

4. A method for removing salt spray interference in circuit board visual abnormal state detection according to claim 3, characterized in that: In the fourth step, the first convolutional layer module first performs convolution on the feature map A, the second convolutional layer module then performs convolution on the feature map A, and then the third convolutional layer module performs convolution on the feature map A. The first convolutional layer module first performs convolution on the feature map A, the second convolutional layer module with more channels then performs convolution on the feature map A, and then the third convolutional layer module with more channels performs convolution on the feature map A.

5. A method for removing salt spray interference in circuit board visual abnormal state detection according to claim 4, characterized in that: The first convolutional layer module first performs convolution on the feature map A, and then the second convolutional layer module with more channels performs convolution on the feature map A, and then the third convolutional layer module with more channels performs convolution on the feature map A. The first convolutional layer module with sixty-four channels first performs convolution on the feature map A, and then the second convolutional layer module with one hundred and twenty-eight channels performs convolution on the feature map A, and then the third convolutional layer module with two hundred and fifty-six channels performs convolution on the feature map A.

6. A method for removing salt spray interference in circuit board visual abnormal state detection according to claim 5, characterized in that: In the fourth step, the convolution of the feature map A2 by the post-convolution layer module in the convolution module group is the convolution of the feature map A2 by the post-convolution layer module with a channel number of 512 in the convolution module group.

7. The method for removing salt spray interference in circuit board visual abnormal state detection according to claim 1, characterized in that: In the third step, the convolution of the feature map A by the encoder module in the encoder module group is that the multiple encoder modules in the encoder module group sequentially convolve the feature map A.

8. The method for removing salt spray interference in circuit board visual abnormal state detection according to claim 7 is characterized in that: In the third step, the multiple encoder modules in the encoder module group sequentially convolve the feature map A, firstly the first encoder module in the encoder module group convolves the feature map A, then the second encoder module in the encoder module group convolves the feature map A, and then the third encoder module in the encoder module group convolves the feature map A.

9. The method for removing salt spray interference in circuit board visual abnormal state detection according to claim 1, characterized in that: In the fifth step, the output module performs convolution on the feature map A4 to extract the features of the feature map A4, and then the output module generates the feature map A5 according to the aforementioned features. The output convolution module in the output module first performs convolution on the feature map A4 to extract the features of the feature map A4, and then the output convolution module generates the feature map A5. In the fifth step, the output module restores the feature map A5 to a circuit board image, and then the output module outputs the circuit board image as a fully connected layer module in the output module. The feature map A5 is restored to a circuit board image, and then the fully connected layer module outputs the circuit board image.

10. The method for removing salt spray interference in circuit board visual abnormal state detection according to claim 1, characterized in that: In the fifth step, the feature map A5 is generated by the output convolution module, and the feature map A5 is first generated by the output convolution module, and then the pooling layer module aggregates the feature map A5 to reduce the size and calculation amount of the feature map A5.

Images