Improvement method of semantic segmentation model based on progressive feature fusion

Abstract

The invention discloses an improvement method of a semantic segmentation model based on progressive feature fusion, and provides a progressive feature fusion method combined with a double-attention mechanism to improve the capability of a decoding module of the semantic segmentation model based on deep learning to recover detail and spatial information so as to obtain better segmentation performance. The method comprises the following steps: (1) designing a double-attention module; (2) designing a progressive feature fusion module; and (3) fusing the progressive feature fusion module into the semantic segmentation model.

Description

[0001] Field [0002] The invention relates to an improved method of a semantic segmentation model based on progressive feature fusion, and belongs to the technical field of semantic segmentation based on deep learning. Background technique [0003] Semantic segmentation of images is another basic task in computer vision besides classification and target detection. It means dividing images into different blocks according to their content, that is, classifying pixels. Compared with classification and detection tasks, Segmentation is finer. [0004] The semantic segmentation model is based on the classification model, which uses convolutional neural network for feature extraction. The high-level abstract features of the image are extracted through repeated pooling and downsampling operations, which are rich in semantics, but at the same time, the resolution of the input image will be gradually reduced, which will inevitably lead to the loss of detailed information, making the s...

AI Technical Summary

Problems solved by technology

The high-level abstract features of the image are extracted through repeated pooling and down-sampling operations, which is rich in semantics, but at the same time it will gradually reduce the resolution of the input image, which will inevitably lead to the loss of detailed information, making the semantic segmentation results too rough.

However, the downsampling layer is very important to improve the receptive field, and it is also essential in the feature extraction process. If the image is kept in the original size during the convolution process, it will bring a very large amount of calculation, which is difficult to apply in practice.

Images