Semantic segmentation method based on improved PSPNet

Abstract

The invention belongs to the technical field of computer vision, and particularly relates to an improved PSPNet-based semantic segmentation method, a lightweight MobileNetV2 network is used for learning feature information, a context semantic feature supplementing module is introduced to reserve more feature, a level set method is introduced as network post-processing after classification prediction is performed on each pixel point in an image in a network, so that a segmentation result is closer to a real contour of a target, and finally, the purpose of semantic segmentation is achieved. Thenetwork model is high in robustness in image semantic segmentation, and segmentation errors are reduced; the MobileNetV2 network with 17 anti-residual units is used as a front-end network of + PSPNet,so that the whole network tends to be lightweight; a context semantic feature supplementing module is introduced, and on the original basis, the problems of feature loss, fuzziness and the like in the sampling process are solved by means of an attention mechanism; and a level set method is introduced into the whole model as a post-processing mode, so that the segmentation precision of the whole model is improved.

Description

technical field [0001] The invention belongs to the technical field of computer vision, in particular to a semantic segmentation method based on improved PSPNet. Background technique [0002] Semantic segmentation covers many scientific fields such as machine learning, computer vision, image processing and human-computer interaction, and has broad application prospects and huge economic value. With the rapid development of artificial intelligence and deep learning, breakthroughs have been made in human body semantic segmentation. [0003] Semantic segmentation refers to the use of some algorithms to allow computers to automatically segment different categories in images. In recent years, there have been many semantic segmentation methods of different deep learning frameworks, including fully convolutional networks (FullyConvolutional Networks, FCN), fully symmetric convolutional networks (UNet), and methods such as DeepLabV3 based on spatial pyramids for image semantics. S...

AI Technical Summary

Problems solved by technology

[0004] The performance of the semantic segmentation method based on deep learning has been greatly improved compared with the traditional segmentation method, but there are still some problems: 1. The problem of inaccurate segmentation edges, the adjacent pixels corresponding to the image information in the receptive field are too similar , if the pixel features of different classes are similar, then the segmentation effect will be greatly affected

2. The network structure used by the traditional PSPNet is a ResNet network, which is suitable for deep networks, but it is not suitable for lightweight networks due to problems such as excessive parameters

3. The traditional semantic segmentation model will lead to problems such as feature blur and feature loss during the feature extraction process, which will affect the experimental results

Images