CN111415325A - A method for defect detection of copper foil substrate based on convolutional neural network - Google Patents

Abstract

The invention discloses a method for detecting defects of a copper foil substrate based on a convolutional neural network, comprising the following steps: collecting and marking a data set; performing data expansion on an image of the data set; constructing a fast and accurate convolutional neural network model; The sample images of the dataset are input into the above-mentioned convolutional neural network model for iterative training to obtain the best detection model; the image of the copper foil substrate to be detected is input into the above-mentioned detection model to identify the image category and realize online automatic detection. The above method is iteratively trained by inputting the data set samples into the constructed convolutional neural network model, thereby obtaining a deep learning detection model, realizing online detection of defective products of copper foil substrates, overcoming the shortcomings of manual design of defect features, improving production efficiency, and thus rapidly Accurately carry out classification detection, has strong adaptability and robustness, and ensures the quality of copper foil substrate products.

Description

A method for defect detection of copper foil substrate based on convolutional neural network

technical field

The invention relates to a copper foil substrate defect detection method, in particular to a copper foil substrate defect detection method based on a convolutional neural network.

Background technique

Some data show that the rapid and accurate detection of copper foil substrate defects is an important research content in industrial production. In the production process of copper foil substrates, it is difficult to avoid appearance defects, which have a great negative impact on the performance and quality of copper foil substrates. , including geometric features, color features, texture features, etc. This detection method has limitations and the implementation process is time-consuming and labor-intensive, and the accuracy and speed are difficult to meet the requirements.

SUMMARY OF THE INVENTION

The invention mainly solves the time-consuming and laborious technical problem of the original manual design feature detection, and provides a copper foil substrate defect detection method based on a convolutional neural network. Obtaining a deep learning detection model, realizing online detection of defective products on copper foil substrates, overcoming the shortcomings of artificially designed defect features, and improving production efficiency, so as to quickly and accurately classify and detect, with strong adaptability and robustness, ensuring copper The quality of foil substrate products.

The above-mentioned technical problems of the present invention are mainly solved by the following technical solutions: the present invention comprises the following steps:

(1) Data set collection and labeling: Collect several types of copper foil substrate defect sample images, classify and label, and collect a class of normal samples for labeling, and use the collected sample images as a data set.

(2) Data expansion is performed on the sample images of the dataset;

(3) Build a convolutional neural network model; the input image size is 96×96×1.

(4) Input the sample image into the above-mentioned convolutional neural network model for iterative training; to obtain the best model.

(5) Input the image of the copper foil substrate to be detected into the above detection model to identify the image category and realize online automatic detection. Realize online automatic detection of defective products of copper foil substrate, thereby improving product quality.

Preferably, the step (2) expands the number of samples by flipping and denoising all the sample images in the data set, and divides the training set and the validation set in a ratio of 9:1.

Preferably, the step (3) convolutional neural network model includes fourteen layers, the first layer is a convolution layer; the second layer is an overlapping maximum pooling layer; the third layer, the fourth layer, and the fifth layer And the seventh layer is the convolution module DepthwiseFire depth separable module of parallel structure; the sixth and eighth layers are the maximum pooling layers; the ninth, tenth, eleventh and twelfth layers are convolutional layers The module DepthwiseResidual is a deeply separable residual module; the thirteenth layer is an average pooling layer, and the fourteenth layer is a softmax classification layer. The softmax classification layer is used to calculate the probability that the output belongs to each class.

Preferably, the first layer of the convolutional neural network model in the step (3) is a convolution layer, including 32 convolution kernels with a receptive field size of 3×3, a stride of 2, and an output of 32 channels and a size of 3. is a 48×48 feature map.

Preferably, the second layer of the convolutional neural network model in the step (3) is an overlapping maximum pooling layer, which is a convolution kernel with a receptive field size of 3×3, a stride of 2, and an output of 32 channels. Feature maps of size 24×24.

Preferably, the third layer, the fourth layer, the fifth layer and the seventh layer of the step (3) convolutional neural network model are the convolution module DepthwiseFire depth separable module of parallel structure, and the third layer is 8 A convolution kernel with a receptive field size of 1×1 and parallel left and right branches, the convolution kernel has a stride of 1, and the output is a feature map with 8 channels and a size of 24×24. The volume of the left branch The product layer is 32 convolution kernels with a receptive field size of 1×1 and a stride of 1. The right branch contains two cascaded convolutional layers, of which the upper convolutional layer has 32 receptive fields with a size of 3×3. The depth of the separable convolution kernel, the stride is 1, the lower convolution layer is 32 convolution kernels with a field size of 1 × 1, the stride is 1, the outputs of the left branch and the right branch are spliced, and the third The final output of the layer is a feature map with 64 channels and a size of 24 × 24; the fourth layer is followed by 12 convolution kernels with a receptive field size of 1 × 1 and parallel left and right branches. The length is 1, and the output is a feature map with 12 channels and a size of 24×24. The convolutional layer of the left branch is 48 convolution kernels with a receptive field size of 1×1, and the stride is 1. The right branch contains two A cascaded convolutional layer, in which the upper convolutional layer is 48 depthwise separable convolution kernels with a receiving field size of 3×3, and the stride is 1, and the lower convolutional layer is 48 receiving fields with a size of 1×1. The convolution kernel, the stride is 1, the output of the left branch and the right branch is spliced, the fourth layer finally outputs a feature map with 96 channels and a size of 24 × 24; the fifth layer is 16 receptive field sizes in turn is a 1×1 convolution kernel and parallel left and right branches. The convolution kernel has a stride of 1, and the output is a feature map with 16 channels and a size of 24×24. The convolution layer of the left branch is 64 A convolution kernel with a receptive field size of 1×1 and a stride of 1. The right branch contains two cascaded convolutional layers, of which the upper convolutional layer is 64 depthwise separable with a receptive field size of 3×3. The convolution kernel, the stride is 1, the lower convolution layer is 64 convolution kernels with a field size of 1×1, the stride is 1, the outputs of the left branch and the right branch are spliced, and the final output of the fifth layer is 128 channels and feature maps of size 24×24; the seventh layer is followed by 24 convolution kernels with a receptive field size of 1×1 and parallel left and right branches. The convolution kernel has a stride of 1 and outputs is a feature map with 24 channels and a size of 12 × 12. The convolutional layer of the left branch is 96 convolution kernels with a receptive field size of 1 × 1, and the stride is 1. The right branch contains two cascaded Convolutional layer, in which the upper convolutional layer is 96 depthwise separable convolution kernels with a receiving field size of 3×3, the stride is 1, and the lower convolutional layer is 96 convolution kernels with a receiving field size of 1×1 , the stride is 1, the outputs of the left branch and the right branch are spliced, and the final output of the seventh layer is a feature map with 192 channels and a size of 12 × 12. The convolution kernel plays a role in compression, and a convolution layer in the left branch and two convolution layers in a cascade structure in the right branch play an expansion role.

Preferably, the sixth and eighth layers of the convolutional neural network model in the step (3) are maximum pooling layers, wherein the sixth layer is a convolutional layer with a receptive field size of 3×3 and a step size of 2 , the output of this layer is a feature map with 128 channels and a size of 12 × 12; the eighth layer is a convolutional layer with a receptive field size of 3 × 3, the stride is 2, and the output of this layer is 192 channels with a size of 6 × 6 feature map.

Preferably, the ninth layer, tenth layer, eleventh layer and twelfth layer of the convolutional neural network model in the step (3) are the convolution module DepthwiseResidual depth separable residual module, wherein the ninth layer and The tenth layer consists of two upper-level convolutional layers plus a lower-level convolutional layer. The upper-level convolutional layer is 256 depthwise separable convolution kernels with a receiving field size of 3×3, and the stride is 1; , the output is a feature map with 256 channels and a size of 6 × 6; the eleventh and twelfth layers are composed of two upper-level convolutional layers and lower-level convolutional layers. The upper-level convolutional layer is 512 depthwise separable convolution kernels with a receiving field size of 3×3, and the stride is 1; , the output is a 512-channel feature map of size 6×6.

Preferably, the thirteenth layer of the convolutional neural network model in step (3) is an average pooling layer, which is a convolutional layer with a receptive field size of 6×6, a step size of 1, and the output of this layer is 512 channels , a feature map of size 1×1.

Preferably, in the step (4), the number of iteration cycles is set to 400, and the validation set accuracy is output once for each iteration cycle. The parameters are fine-tuned or the number of iteration cycles is modified according to the accuracy verification to obtain the optimal model.

The beneficial effects of the present invention are: by inputting the data set samples into the iterative training of the convolutional neural network model constructed, a deep learning detection model is obtained, online detection of defective copper foil substrate products is realized, the shortcomings of artificial design of defect features are overcome, and the Production efficiency, so as to quickly and accurately classify and detect, with strong adaptability and robustness, ensuring the quality of copper foil substrate products.

Detailed ways

The technical solutions of the present invention are further described in detail below through the examples.

Embodiment: A method for detecting defects of copper foil substrates based on convolutional neural network in this embodiment includes the following steps:

(1) Dataset collection and labeling. Several types of copper foil substrate defect sample images are collected, classified and marked, and one type of normal samples are collected and marked, and the collected above-mentioned sample images are used as a data set.

(2) Data expansion is performed on the sample images of the dataset. All sample images in the dataset are expanded by flipping and denoising, and all samples are divided into training set and validation set in a ratio of 9:1.

(3) Build a convolutional neural network model with an input image size of 96×96×1.

The first layer is a convolutional layer, including 32 convolution kernels with a receptive field size of 3 × 3, a stride of 2, and the output is a feature map with 32 channels and a size of 48 × 48.

The second layer is the overlapped max pooling layer, which is a convolution kernel with a receptive field size of 3×3, a stride of 2, and the output is a 32-channel feature map of size 24×24.

The third layer adopts the convolution module DepthwiseFire depth separable module in parallel structure. The first is 8 convolution kernels with a receptive field size of 1 × 1, stride 1, and the output is a feature map with 8 channels and a size of 24 × 24, which plays a role in compression. Then there are the parallel left and right branches, in which the convolutional layer of the left branch is 32 convolution kernels with a receptive field size of 1×1, and the stride is 1, which plays an expansion role. The right branch consists of two cascaded convolutional layers. The upper-level convolutional layer has 32 depthwise separable convolution kernels with a receiving field size of 3×3, and the stride is 1; the lower-level convolutional layer has 32 receiving fields with a size of 1. The convolution kernel of ×1, with a stride of 1, plays an expansion role. Finally, the outputs of the left and right branches are spliced, and the output of this layer is a feature map with 64 channels and a size of 24×24.

The fourth layer also adopts the convolution module DepthwiseFire depth separable module in parallel structure. The first is 12 convolution kernels with a receptive field size of 1 × 1, stride 1, and the output is a feature map with 12 channels and a size of 24 × 24, which plays a role in compression. Then there are the parallel left and right branches, in which the convolutional layer of the left branch is 48 convolution kernels with a receptive field size of 1×1, and the stride is 1, which plays an expansion role. The right branch consists of two cascaded convolutional layers. The upper-level convolutional layer has 48 depthwise separable convolution kernels with a receiving field size of 3 × 3 and a stride of 1; the lower-level convolutional layer has 48 receiving fields with a size of 1. The convolution kernel of ×1, with a stride of 1, plays an expansion role. Finally, the outputs of the left and right branches are spliced, and the output of this layer is a feature map with 96 channels and a size of 24×24.

The fifth layer also adopts the convolution module DepthwiseFire depth separable module in parallel structure. The first is 16 convolution kernels with a receptive field size of 1 × 1, stride 1, and the output is a feature map with 16 channels and a size of 24 × 24, which plays a role in compression. Then there are the parallel left branch and right branch, where the convolutional layer of the left branch is 64 convolution kernels with a receptive field size of 1×1, and the stride is 1, which plays an expansion role. The right branch consists of two cascaded convolutional layers. The upper-level convolutional layer has 64 depthwise separable convolution kernels with a receiving field size of 3×3, and the stride is 1; the lower-level convolutional layer has 64 receiving fields with a size of 1. The convolution kernel of ×1, with a stride of 1, plays an expansion role. Finally, the outputs of the left and right branches are spliced, and the output of this layer is a feature map with 128 channels and a size of 24×24.

The sixth layer adopts the maximum pooling layer, which is a convolutional layer with a receptive field size of 3 × 3 and a stride of 2. The output of this layer is a feature map with 128 channels and a size of 12 × 12.

The seventh layer also adopts the convolution module DepthwiseFire depth separable module in parallel structure. The first is 24 convolution kernels with a receptive field size of 1 × 1, stride 1, and the output is a feature map with 24 channels and a size of 12 × 12, which plays a role in compression. Then there are the parallel left branch and right branch, where the convolution layer of the left branch is 96 convolution kernels with a receptive field size of 1 × 1, and the stride is 1, which plays an expansion role. The right branch consists of two cascaded convolutional layers. The upper-level convolutional layer has 96 depthwise separable convolution kernels with a receiving field size of 3×3, and the stride is 1; the lower-level convolutional layer has 96 receiving fields with a size of 1. The convolution kernel of ×1, with a stride of 1, plays an expansion role. Finally, the outputs of the left and right branches are spliced, and the output of this layer is a feature map with 192 channels and a size of 12×12.

The eighth layer adopts the max pooling layer, which is a convolutional layer with a receptive field size of 3 × 3 and a stride of 2. The output of this layer is a feature map of 192 channels and a size of 6 × 6.

The ninth layer adopts the convolution module DepthwiseResidual depthwise separable residual module. It consists of two upper-level convolutional layers and a lower-level convolutional layer. The upper-level convolutional layer is 256 depthwise separable convolution kernels with a receiving field size of 3×3, and the stride is 1; , the output is a 256-channel feature map of size 6×6.

The tenth layer also adopts the convolution module Depthwise Residual depth separable residual module. It consists of two upper-level convolutional layers and a lower-level convolutional layer. The upper-level convolutional layer is 256 depthwise separable convolution kernels with a receiving field size of 3×3, and the stride is 1; , the output is a 256-channel feature map of size 6×6.

The eleventh layer adopts the convolution module DepthwiseResidual depthwise separable residual module. It consists of two upper-level convolutional layers and a lower-level convolutional layer. The upper-level convolutional layer is 512 depthwise separable convolution kernels with a receiving field size of 3×3, and the stride is 1; , the output is a 512-channel feature map of size 6×6.

The twelfth layer adopts the convolution module Depthwise Residual depth separable residual module. It consists of two upper-level convolutional layers and a lower-level convolutional layer. The upper-level convolutional layer is 512 depthwise separable convolution kernels with a receiving field size of 3×3, and the stride is 1; , the output is a 512-channel feature map of size 6×6.

The thirteenth layer adopts an average pooling layer, which is a convolutional layer with a receptive field size of 6×6 and a stride of 1. The output of this layer is a feature map with 512 channels and a size of 1×1.

The fourteenth layer is the softmax classification layer, which is used to calculate the probability that the output belongs to each class.

(4) Input the sample image into the above convolutional neural network model for iterative training. Set the number of iteration cycles to 400, and output the validation set accuracy once for each iteration cycle. The parameters can be fine-tuned or the number of iteration cycles can be modified to obtain the optimal model.

(5) Input the image of the copper foil substrate to be detected into the above detection model to identify the image category, and realize the online automatic detection of defective products of the copper foil substrate, thereby improving work efficiency and ensuring product quality.

Claims (10)

1. a copper foil substrate defect detection method based on convolutional neural network, is characterized in that, comprises the following steps: (1) Data set collection and labeling; (2) Data expansion is performed on the sample images of the dataset; (3) Build a convolutional neural network model; (4) Input the sample image into the above-mentioned convolutional neural network model for iterative training; (5) Input the image of the copper foil substrate to be detected into the above detection model to identify the image category and realize online automatic detection. 2 . The method for detecting defects of copper foil substrates based on convolutional neural network according to claim 1 , wherein the step (2) expands the samples by flipping and denoising all the sample images in the data set. 3 . The training set and the validation set are divided according to the ratio of 9:1. 3. The method for detecting defects of copper foil substrates based on convolutional neural network according to claim 1, wherein the step (3) convolutional neural network model comprises fourteen layers, and the first layer is a convolutional neural network. layer; the second layer is an overlapping max pooling layer; the third, fourth, fifth and seventh layers are convolutional modules with parallel structure DepthwiseFire depth separable modules; the sixth and eighth layers are the maximum Pooling layer; the ninth, tenth, eleventh and twelfth layers are convolution modules Depthwise Residual depth separable residual module; the thirteenth layer is the average pooling layer, and the fourteenth layer is the softmax classification Floor. 4. The method for detecting defects of copper foil substrates based on convolutional neural network according to claim 3, wherein the first layer of the convolutional neural network model in the step (3) is a convolutional layer, comprising 32 A convolution kernel with a receptive field size of 3 × 3, a stride of 2, and the output is a feature map with 32 channels and a size of 48 × 48. 5 . The method for detecting defects of copper foil substrates based on convolutional neural network according to claim 3 , wherein the second layer of the convolutional neural network model in the step (3) is an overlapping maximum pooling layer. 6 . , is a convolution kernel with a receptive field size of 3 × 3, a stride of 2, and the output is a feature map with 32 channels and a size of 24 × 24. 6 . The method for detecting defects of copper foil substrates based on convolutional neural network according to claim 3 , wherein the step (3) of the third layer, the fourth layer and the fifth layer of the convolutional neural network model is as follows. 7 . The layer and the seventh layer are the convolution module DepthwiseFire depth separable module with parallel structure. The third layer is followed by 8 convolution kernels with a receptive field size of 1×1 and parallel left and right branches. The convolution kernel Step size 1, the output is a feature map with 8 channels and a size of 24×24, the convolution layer of the left branch is 32 convolution kernels with a receptive field size of 1×1, the step size is 1, and the right branch contains Two cascaded convolutional layers, in which the upper convolutional layer is 32 depthwise separable convolution kernels with a field size of 3×3, and the stride is 1, and the lower convolutional layer is 32 with a field size of 1× The convolution kernel of 1, the stride is 1, the outputs of the left branch and the right branch are spliced, and the final output of the third layer is a feature map with 64 channels and a size of 24 × 24; the fourth layer is followed by 12 receptive fields A convolution kernel with a size of 1×1 and parallel left and right branches, the convolution kernel has a stride of 1, and the output is a feature map with 12 channels and a size of 24×24. The convolutional layer of the left branch is 48 convolution kernels with a receptive field size of 1×1 and a stride of 1. The right branch contains two cascaded convolutional layers, of which the upper-level convolutional layer is 48 depth-receptive layers with a receptive field size of 3×3. Separate the convolution kernel, the stride is 1, the lower convolution layer is 48 convolution kernels with a field size of 1×1, the stride is 1, the outputs of the left branch and the right branch are spliced, and the fourth layer is the final output It is a feature map with 96 channels and a size of 24 × 24; the fifth layer is followed by 16 convolution kernels with a receptive field size of 1 × 1 and parallel left and right branches. The step size of the convolution kernel is 1, The output is a feature map with 16 channels and a size of 24 × 24. The convolutional layer of the left branch is 64 convolution kernels with a receptive field size of 1 × 1, and the stride is 1. The right branch contains two cascades The upper-level convolutional layer is 64 depthwise separable convolution kernels with a field size of 3×3, and the stride is 1, and the lower-level convolutional layer is 64 convolutional layers with a field size of 1×1. The kernel, with a stride of 1, splices the outputs of the left branch and the right branch. The final output of the fifth layer is a feature map with 128 channels and a size of 24×24; the seventh layer is followed by 24 receptive fields with a size of 1× 1 convolution kernel and parallel left and right branches, the convolution kernel has a stride of 1, and the output is a 24-channel feature map with a size of 12 × 12, and the convolutional layer of the left branch is 96 receptive fields A convolution kernel of size 1×1 with a stride of 1, the right branch contains two concatenated convolutional layers, where the upper convolutional layer is 96 depthwise separable convolution kernels with a field size of 3×3 , the stride is 1, the lower convolutional layer is 96 convolution kernels with a field size of 1×1, the stride is 1, the outputs of the left branch and the right branch are spliced, and the final output of the seventh layer is 192 channels , a feature map of size 12 × 12. 7. a kind of copper foil substrate defect detection method based on convolutional neural network according to claim 3, is characterized in that, the sixth layer and the eighth layer of described step (3) convolutional neural network model are maximum pools The sixth layer is a convolutional layer with a receptive field size of 3×3 and a stride of 2. The output of this layer is a feature map of 128 channels and a size of 12×12; the eighth layer is a receptive field size of 3 A ×3 convolutional layer with a stride of 2, this layer outputs a 192-channel feature map of size 6×6. 8 . The method for detecting defects of copper foil substrates based on convolutional neural network according to claim 3 , wherein the step (3) of the ninth layer, the tenth layer and the tenth layer of the convolutional neural network model is as follows. 9 . The first layer and the twelfth layer are the convolution module Depthwise Residual depth separable residual module, of which the ninth and tenth layers are composed of two upper-level convolutional layers and lower-level convolutional layers. The upper-level convolutional layer is 256 depthwise separable convolution kernels with a receiving field size of 3×3, and the stride is 1; , the output is a feature map with 256 channels and a size of 6 × 6; the eleventh and twelfth layers are composed of two upper-level convolutional layers and lower-level convolutional layers. The upper-level convolutional layer is 512 depthwise separable convolution kernels with a receiving field size of 3×3, and the stride is 1; , the output is a 512-channel feature map of size 6×6. 9 . The method for detecting defects of copper foil substrates based on a convolutional neural network according to claim 3 , wherein the thirteenth layer of the convolutional neural network model in the step (3) is an average pooling layer, 10 . is a convolutional layer with a receptive field size of 6×6 and a stride of 1. The output of this layer is a 512-channel feature map of size 1×1. 10 . The method for detecting defects of copper foil substrates based on convolutional neural network according to claim 1 , wherein in the step (4), the number of iteration cycles is set to 400, and a verification set is output for each iteration cycle. 11 . Accuracy.