Apparatus and method for image classification and segmentation based on feature-guided network, device, and medium

An apparatus for image classification and segmentation based on a feature-guided network, characterized by comprising:

• a data preprocessing module, configured to perform various affine transformation operations and normalization on an input image;
• an image classification module, configured to construct a feature-guided classification network, wherein the feature-guided classification network comprises one 3×3 convolutional layer, five basic unit blocks, and a fully-connected network;
• and
• an image segmentation module, configured to construct a feature-guided segmentation network, wherein the feature-guided segmentation network adopts an encoder-decoder structure;
• the encoder comprises one 3×3 convolutional layer, four basic unit blocks and downsampling layers, and one basic unit block and a rectified linear unit (ReLU) activation function;
• and the decoder comprises five upsampling layers and five regular convolutional layers,
• wherein whether to classify or segment an input image based on a requirement of the input image is chosen, and if choosing to classify the image, the image is input to a trained classification network model and outputting a corresponding category;
• or if choosing to segment the image, the image is input to a trained segmentation network model and outputting a segmented image;
• in the feature-guided classification network, the input image is processed by one 3×3 convolutional layer, and then the five basic unit blocks and downsampling layers, and finally, an obtained output feature map is flattened and connected to the fully-connected network;
• residual connections are added before and after each basic unit block;
• and a local feature is enhanced in a spatial dimension and a global feature is extracted in a channel dimension for an output feature map generated by each basic unit block, and the global feature and local feature are fused before a fully-connected layer;
• in the encoder, the input image is processed by one 3×3 convolution layer, then four basic unit blocks and downsampling layers, and finally one basic unit block and a rectified linear unit (ReLU) activation function;
• residual connections are added before and after each basic unit block;
• the local feature is enhanced in the spatial dimension and the global feature is extracted in the channel dimension for an output feature map generated by each basic unit block;
• and in the decoder, an output feature map generated by the ReLU activation function is input into the decoder, the output feature map generated by each basic unit block in the encoder and a feature map generated by each upsampling layer in the decoder with a corresponding scale are connected in the channel dimension, then regular convolution is performed, a generated output feature map is used as an input feature map of a next upsampling layer, and a segmented image is output until fifth upsampling and fifth regular convolution are performed on an output feature map generated after fourth regular convolution.