CN115100393A

CN115100393A - Deep learning-based PCB welding defect detection method

Info

Publication number: CN115100393A
Application number: CN202210722138.7A
Authority: CN
Inventors: 陈树越; 王佳宏
Original assignee: Changzhou University
Current assignee: Changzhou University
Priority date: 2022-06-24
Filing date: 2022-06-24
Publication date: 2022-09-23

Abstract

The invention relates to the technical field of defect detection, in particular to a PCB welding defect detection method based on deep learning, which comprises the steps of performing augmentation processing on a PCB image in a data set, completing conversion of data format and size, and dividing the data set; determining the number of regression prior frames by using a K-Means algorithm, and determining the size of an anchor frame by using a scaling formula; adjusting the receptive field of the NECK area, and adding an SE module when the downsampling of the Backbone area is completed; and inputting the converted data file into a training function for training, and setting training parameters. The invention solves the problems of low efficiency and low precision of the current PCB defect detection, continuously enhances the defect recognition capability of a machine through deep learning training, determines the size of an anchor frame by utilizing a K-Means algorithm, and accurately positions the defect part through an anchor frame scaling formula, thereby improving the defect detection rate and preventing the poor circuit board from flowing out.

Description

Deep learning-based PCB welding defect detection method

Technical Field

The invention relates to the technical field of defect detection, in particular to a PCB welding defect detection method based on deep learning.

Background

Printed Circuit Boards (PCBs) are essential components of all electronic devices and have evolved over the years into a complex assembly. The defects of missing holes, mouse bites, open circuits, short circuits, open circuits, stray, false copper and the like exist in the production process of the current PCB, the appearance of the PCB is influenced, and the use performance of the PCB can be damaged. The traditional PCB image to be detected is compared with a pre-specified standard template in the detection process based on machine vision detection defects, the detection flow is complex and is easily influenced by the environment, and the detection speed and the detection precision cannot be adapted to a quick production link. The target detection method based on deep learning can be simply divided into a two-stage algorithm represented by an R-CNN series and a one-stage algorithm represented by an SSD or a YOLO series, and the defect detection problem is converted into a classification regression problem. The PCB defect detection based on deep learning is that learning training is carried out through a mass data set, a defect rule is searched from the defect rule and parameters are saved, when a target to be detected is changed, only a new data set is needed to be added for learning training, and compared with a traditional machine learning method, the method based on deep learning is stronger in adaptability, higher in detection speed and precision and capable of adapting to a quick production detection link.

Disclosure of Invention

Aiming at the defects of the existing algorithm, the invention solves the problems of low efficiency and low precision of the current PCB defect detection, continuously enhances the defect recognition capability of the machine through deep learning training, utilizes the K-Means algorithm to determine the size of the anchor frame, and precisely positions the defect part through the anchor frame scaling formula, thereby improving the defect detection rate and preventing the poor circuit board from flowing out.

The technical scheme adopted by the invention is as follows: a PCB welding defect detection method based on deep learning comprises the following steps:

s1, performing amplification processing on the PCB image in the data set, completing conversion of data format and size, and dividing the data set;

s2, determining the number of regression prior frames by using a K-Means algorithm, and determining the size of an anchor frame by using a scaling formula;

further, the anchor frame size is recalculated using a 1-IOU based K-means algorithm to generate 9 sets of anchor frames, 12,31 13,15 16,26 20,13 33,66 38,48 45,54 56,34 81, 87.

Further, the scaling formula is:

x′ ₁ ＝Ax ₁ (1)

in the formula, x _i The width of the ith group of anchor frames (in the order of the width size of the anchor frames from small to large), i is 2,3.. 9; x' _i The width of the zoomed anchor frame; a is the reduction multiple of the anchor frame, and A is 0.65; y is _i The height of the ith group of anchor frames; y' _i Is the zoomed anchor frame height.

S3, adjusting the receptive field of the NECK area, and adding an SE module when the backhaul area finishes downsampling;

further, the method comprises the steps of deleting 19 × 19 characteristic branches of the receptor field of the NECK region and adding 152 × 152 characteristic branches;

s4, inputting the converted data file into an improved Yolov5S network for training, and setting training parameters;

further, the iteration number epoch is set to 200, and the img-size is set to 608 × 608.

The invention has the beneficial effects that:

1. the method can quickly learn the characteristics of the tiny defects, and solves the problems of low efficiency and low accuracy of the traditional defect detection; the method for clustering Anchors in the K-Means algorithm is set to be based on the 1-IOU distance, and scaling processing is carried out according to the defect types, so that the defect types can be better covered.

Drawings

FIG. 1 is a PCB defect detection flow chart based on deep learning of the present invention;

FIG. 2 is a graph of the detection effect of the method of the present invention in FIG. 1;

FIG. 3 is a graph of the detection effect of the method of the present invention FIG. 2;

FIG. 4 is a graph of the detection effect of the method of the present invention shown in FIG. 3.

Detailed Description

The invention will be further described with reference to the accompanying drawings and examples, which are simplified schematic drawings and illustrate only the basic structure of the invention in a schematic manner, and therefore only show the structures relevant to the invention.

As shown in fig. 1, a PCB welding defect detection method based on deep learning includes the following steps:

s1, performing amplification processing on the PCB image in the data set to complete conversion of data format and size, and dividing the data set;

performing amplification processing on the PCB image in the data set, wherein the amplification processing comprises cutting, splicing and rotating, labeling part of the defects of the PCB image by adopting Labelme software to obtain a generated json format label file, wherein the label file comprises six common defects and defect category names of the PCB, converting the xml file in the original data set and the json file in the amplification part into a txt file, and enabling the image and the label to be as follows: 2, dividing the training set into an original training set and an original testing set in proportion;

the method is characterized in that a single picture in the original data set is randomly combined corresponding to a single defect to form multiple defects of the picture, and meanwhile, the method for changing the brightness and adding Gaussian noise is adopted, so that the value of a boundary frame is not changed, different environments under various actual production can be met, and the overfitting risk of a neural network during training is reduced;

the defect images of the PCBs in the data set are captured by conventional industrial cameras, and usually have a larger resolution, such as 3034 x 1586 and 2544 x 2516, and the defect images are smaller in proportion than the whole image, so that all the image sizes are set to 608 x 608 in the training function, and the determination of the anchor frame is facilitated.

S2, inputting the converted data file into a Yolov5S network, wherein the Yolov5S network mainly comprises an input end, a backsbone, a Neck and a preview, and compared with Yolov5m, Yolov5x and Yolov5l, the network has the defects that the speed is higher when a small target is detected, but the precision is not high enough, and the following is an improvement in the network structure:

s21, modifying the self-adaptive generation anchor frame of the input end, because the width and height difference of the anchor frame of the target defect existing in the defect type is too large, and the error of the clustering result is very large by using a Euclidean distance formula, modifying the self-adaptive generation anchor frame of the input end, replacing the sample distance measurement of the K-means algorithm from the Euclidean distance to the IOU distance, calculating 9 groups of anchor frames which are respectively [12,31] [13,15] [16,26] [20,13] [33,66] [38,48] [45,54] [56,34] [81,87], wherein the return value Accuracy is 0.804815, reflecting the IOU values of all the box and other clustering center anchors, and the larger value indicates that the more marked frames can be adapted by K anchors.

Determining a group of scaling formulas according to the actual defect type size to ensure that the scaled anchor frame can more effectively cover the mark frame in the data set, wherein the scaling formulas are as follows:

x′ ₁ ＝Ax ₁ (1)

S3, deleting 19 × 19 characteristic branches of the NECK regional receptive field, and adding 152 × 152 characteristic branches; and an SE module is added when the backup area finishes downsampling, so that the importance degrees of different channel characteristics are automatically learned, and the sensitivity of the model to the channel characteristics is improved.

The NECK area of Yolov5s adopts the structure of FPN + PAN, so that through the combination operation, the FPN layer transmits semantic features from top to bottom, and the feature pyramid transmits the semantic features from bottom to top to be determined as features, thereby further improving the feature extraction capability; the original Yolov5s feature maps were 19 × 19, 38 × 38, and 76 × 76, and since the defect target was too small compared to the overall image, the 19 × 19 feature branch with the largest receptive field in the NECK region was deleted and 152 × 152 feature branches were added.

Compared with other Yolov5 series networks, the Yolov5s has shallow depth and width, so an attention mechanism is added, an SE module is added when a backsbone area completes a down-sampling process, a new characteristic diagram is obtained through learning calculation, and the SE module is added in the original Yolov5s network structure, so that less calculation amount is increased, and performance is improved.

And S4, inputting the converted data file into an improved Yolov5S network in S3 for training, setting training parameters, setting the iteration time epoch to be 200 and the img-size to be 608 × 608, testing the trained model, and performing performance evaluation and optimization.

And the evaluation of the model takes mAP @0.5:0.95 as a standard, the training model is evaluated in performance, and when the evaluation fails to reach the required standard, the iteration times can be adjusted while the data set is enriched. When test set inspection is carried out on a subsequent test function, the conf-thres is set to be 0.5, and the iou-thres is set to be 0.45, so that the problems of overlapping of anchor frames and non-display of partial defects can be effectively avoided; in actual production testing, the value of conf-thres may be increased or the value of iou-thres may be decreased if a test frame overlap is encountered.

The actual measurement results are shown in fig. 2,3 and 4, in which a rectangular frame is used for detecting the defect part, and the defect type and the confidence coefficient are arranged above the rectangular frame.

The actual measurement result shows that the method can effectively detect the defects in the PCB production process.

In light of the foregoing description of the preferred embodiment of the present invention, many modifications and variations will be apparent to those skilled in the art without departing from the spirit and scope of the invention. The technical scope of the present invention is not limited to the content of the specification, and must be determined according to the scope of the claims.

Claims

1. A PCB welding defect detection method based on deep learning is characterized by comprising the following steps:

s2, determining the prior frame number of the regression by using a K-Means algorithm, and determining the size of an anchor frame by a scaling formula;

and S4, inputting the converted data file into an improved Yolov5S network for training, and setting training parameters.

2. The deep learning-based PCB welding defect detection method of claim 1, wherein the prior frame number is based on recalculating the anchor frame size by the K-means algorithm of the 1-IOU to generate 9 sets of anchor frames, respectively [12,31] [13,15] [16,26] [20,13] [33,66] [38,48] [45,54] [56,34] [81,87 ].

3. The deep learning based PCB welding defect detection method of claim 1, wherein the scaling formula is as follows:

x′ ₁ ＝Ax ₁ (1)

4. The deep learning-based PCB welding defect detection method of claim 1, wherein the adjusting of the NECK regional reception field is to delete 19 × 19 characteristic branches of the NECK regional reception field and add a 152 × 152 characteristic branch.

5. The deep learning-based PCB welding defect detection method of claim 1, wherein the training parameters comprise: the number of iterations epoch is 200 and img-size 608.