Semantic segmentation method based on self-attention mechanism and dilated convolution pooling, storage medium and visual device

The invention relates to the technical field of computer vision, in particular to a semantic segmentation method based on a self-attention mechanism and dilated convolution pooling, a storage medium and a vision device.The method comprises the following steps that S1, an encoder is adopted for extracting features of an input image, multiple layers of features are extracted from shallow to deep respectively, and the multiple layers of features are extracted; the feature size of the multi-layer feature is halved layer by layer in sequence, and the number of channels of the multi-layer feature is increased layer by layer in sequence; s2, a self-attention structure is adopted to solve an autocorrelation matrix of the feature Xi of each layer, matrix multiplication is carried out on the autocorrelation matrix and the original input feature Xi, global self-attention information is obtained, and a feature Y is output; s3, performing receptive field enhancement on the input feature Yi of each layer by adopting a ligh