Image semantic segmentation method

Abstract

The invention discloses an image semantic segmentation method, comprising two parts of encoding and decoding, wherein the encoding end uses a series of convolution operations and a maximum pool operation to extract features based on the classical FCN model, the number of channels of the feature map after convolution is doubled, and the length and width of the feature map after maximum pool are reduced by half; the decoder first upsamples the feature map extracted from the convolutional layer of the encoder, then connects it with the feature map extracted from the double upsampling of the decoder, and then extracts the comprehensive features from the convolutional layer. In this way, the shallow information and the deep information can be merged well, and the network outputs a 21-dimensional matrix with the same size as the original image. The technical proposal and application of the invention redefines the network structure, fully utilizes the context information of the network by combining the decoding end and the characteristic map of the decoding end, and improves the final accuracy to a certain extent. The resolution of the segmented image relative to the original image is preserved.

Description

technical field [0001] The invention relates to an image computer vision analysis and processing method, in particular to a method for realizing image semantic segmentation by using deep learning. Background technique [0002] With the continuous deepening of computer vision research, researchers have gradually turned their attention to more accurate analysis and understanding of images. The problem of semantic segmentation is proposed to meet this requirement. The fundamental purpose of semantic segmentation is to determine the semantic category of each pixel in the image by training the content of the image. The following are some achievements in the field of image semantic segmentation in recent years. [0003] Fully Convolutional Networks (FCN, Fully Convolutional Networks) can be said to be the pioneering work of deep learning in image semantic segmentation tasks. It was proposed by the research team of the University of California, Berkeley, and promoted the original...

AI Technical Summary

Problems solved by technology

One is that continuous convolution and pooling greatly reduce the resolution of feature maps, which is very unfavorable for dense prediction tasks. Although FCN-8s is much better than FCN-32s, the result of upsampling is still relatively rough. Insensitive to details in the image

The second is that FCN classifies each pixel without considering the relationship between pixels and pixels.

[0006] However, EDNs still have the following disadvantages: the upsampling zero padding operation introduces noise; while the network structure of EDNs has been solidified, the skip connection layer only connects the encoding end and the corresponding decoding end, resulting in the context information not being fully utilized

Images