Complex industrial image edge extraction method based on artificial intelligence and complex industrial image edge extraction device thereof

A complex industrial image edge extraction device based on artificial intelligence, comprising:

• an image pyramid processing module;
• an image cropping processing module;
• an image scaling processing module;
• and
• an edge feature extraction module;
• wherein the image pyramid processing module is configured to obtain first image data to be subjected to edge extraction, and perform image pyramid processing on the first image data to be subjected to the edge extraction to obtain second image data;
• the image cropping processing module is configured to perform image cropping processing on the second image data to obtain third image data;
• the image scaling processing module is configured to perform image scaling processing on the third image data to obtain first scaled image data;
• and the edge feature extraction module is configured to input the first scaled image data into an artificial intelligence model for training and convergence and performing edge extraction processing, and then output an edge extraction result of the first scaled image data;
• wherein the artificial intelligence model for training and convergence sequentially comprises convolution mudules, BottleneckCSP network modules/C3 network modules, spatial pyramid pooling-fast (SPPF) modules, up-sampling modules, and Concat modules;
• wherein the edge feature extraction module is configured to input the first scaled image data into the artificial intelligence model for training and convergence and performing the edge extraction processing, the first scaled image data are input into the artificial intelligence model for training and convergence, and the feature extraction processing is performed on the artificial intelligence model for training and convergence through the up-sampling modules to obtain the sampling extraction feature image;
• the sampling extraction feature image is input into the BottleneckCSP network modules/C3 network modules and the feature extraction processing is performed to obtain an extraction feature, then the extraction feature is input into the SPPF modules;
• the extraction feature is copied through the SPPF modules to obtain copy extraction features, the copy extraction features are copied through the Concat modules to obtain connection output feature data;
• the connection output feature data pass through a convolution layer having a convolution kernel number of 32 to generate 32*320*320 output feature data, and the 32*320*320 output feature data are input into a backbone network module through batch normalization (BN) and Leaky Relu;
• convolution processing is performed in the backbone network model sequentially through a 1*1 convolution layer and a 3*3 convolution layer to obtain a convolution result, and the convolution result is added to the sampling extraction feature image through a residual structure to obtain additive feature data;
• the additive feature data are added to an adjacent layer through adaptive feature pooling, and Concat operation is performed in the Concat modules to obtain a Concat operation result;
• regression loss calculation processing and non-maximum suppression algorithm calculation processing are sequentially performed on the Concat result through GOU_Loss to obtain a calculation result, and the calculation result is output as the edge extraction result of the first scaled image data;
• wherein the feature extraction processing is performed on the artificial intelligence model for training and convergence through the up-sampling modules to obtain the sampling extraction feature image, a slice operation is performed on the first scaled image data through the up-sampling modules in the artificial intelligence model for training and convergence to obtain second scaled image data, the feature extraction processing is performed on the second scaled image data through convolution kernels to obtain a feature extraction result;
• the slice operation is performed on the first 320*320*32 feature map through the up-sampling modules to obtain a second 320*320*32 feature map, the feature extraction processing is performed on the second 320*320*32 feature map to obtain a first extraction result, and a convolution operation of 64 convolution kernels is performed on the first extraction result to obtain a 160*160*64 feature map;
• wherein the sampling extraction feature image is input into the BottleneckCSP network modules/C3 network modules and the feature extraction processing is performed to obtain the extraction feature, and the extraction feature is input into the SPPF modules;
• the sampling extraction feature image is input into the BottleneckCSP network modules/C3 network modules, and image feature information including structures, colors, edges, and locations is extracted through summing;
• fusion processing is performed on the image feature information including the structures, the colors, the edges, and the locations to obtain a fused image feature information;
• and the feature extraction processing is performed on the fused image feature information based on common convolution kernels to obtain a second extraction result, up-sampling processing is performed on the second extraction result to obtain the extraction feature, and the extraction feature is input into the SPPF modules.