Why are encoder and decoder structures often used in image segmentation tasks?

Image segmentation

Encoder

The encoder can be regarded as a feature extraction network , Pooling layers are often used to gradually reduce the size of the input data

decoder

The decoder uses up sampling / Deconvolution and other network layers to gradually restore the details of the target and the corresponding spatial dimensions .

Introducing pool layer can increase receptive field , And it can make feature extraction focus on important information , Reduce background interference , It is helpful for image classification . However, the pooling operation causes a large loss of location information , The features extracted by the encoder are not enough to segment the pixels accurately , This decoder caused great trouble , Make the decoder sample directly from the top / The segmentation image generated by deconvolution is rough . therefore , Some researchers propose to establish a fast connection between encoder and decoder , The high-resolution feature information can participate in the subsequent decoding process , And then help the decoder better recover the details of the target .