Image Semantic Segmentation Based On Context Information And Anisotropic Pooling

Image is a kind of information carrier to describe the objective world.It is an important way for us to recognize the world,obtain information,and transfer value.It can help us better perceive the world and make judgments.Research in the field of images has never stopped.Especially with the development of big data,artificial intelligence,algorithms and computing power,image segmentation methods based on deep learning have made great breakthrough.Image semantic segmentation aims to achieve pixel-level understanding by assigning a specific category label to each pixel point in the input image.As a key technology in the field of computer vision,semantic segmentation has also been implemented,providing reliable solutions in automatic driving,medical diagnosis and significant object detection.This paper aims to explore the problems in the field of image semantic segmentation based on deep learning.In semantic segmentation task,the similar appearance between different categories of objects and the different size of the same category of objects bring difficulties to pixel-level segmentation,and the fuzzy segmentation and incorrect segmentation of the edge details of objects also aggravate the uncertainty.For the above problems,this paper combined with a variety of technology,using multi-scale fusion,feature enhancement,anisotropic pooling,attention mechanism and auxiliary supervision methods such as strategy,in a more flexible and efficient ways to capture the spatial characteristics and all kinds of context information,effectively and successfully improve the segmentation performance of the network.We summarize the main work and contribution as follows:(1)Aiming at the problem of low segmentation accuracy due to lack of useful context information in semantic segmentation task,this paper proposes a bidirectional context hybrid semantic segmentation network based on multi-level aggregation.Combining with the characteristics of attention mechanism,this method uses feature enhancement module,multi-level context aggregation module and bidirectional context mixing module to capture various useful context information.(2)Aiming at the problem of fuzzy segmentation and missegmentation of object edge details,this paper proposes a non-local attention semantic segmentation network based on anisotropic pooling.In this method,an anisotropic pooling module,which is different from the traditional square pooling,is designed to deeply encode the context features,and the interdependence between feature channels is modeled by non-local channel context to optimize the segmentation effect of object edge details.(3)Using Python language and Pytorch deep learning framework,a large number of experiments are carried out on PASCAL VOC 2012 dataset,Cam Vid dataset and Cityscapes dataset.The experimental results demonstrate that the two algorithms proposed in this paper can effectively and successfully improve the segmentation performance.