Research On Lightweight Dual-resolution Network For Semantic Segmentation Based On Boundary Assistance

Image semantic segmentation is a fundamental and challenging task in the computer vision,which plays a significant role in some real-world applications,such as robot sensing,autonomous driving,and so on.The goal of semantic segmentation is to assign a unique semantic category label to each pixel in image.With the development of convolutional neural networks(CNNs),some accurate networks have been proposed for semantic segmentation,which have hundreds even thousands of convolutional layers and feature channels.Although these networks have achieved impressive results of semantic segmentation,their complicated network architecture and heavy computational burden make them difficult to deploy into resource-constrained edge devices.Thus,many researchers prefer to design lightweight semantic segmentation networks.To compress model size,the existing mainstream lightweight semantic segmentation networks simplify network architecture,resulting in inadequate feature extraction and limiting the performance of semantic segmentation.Moreover,most of these networks neglect the benefit of boundary information,resulting in the constructed model having insufficient capacity of boundary expression and making the final prediction around object boundary unclear.To deal with these shortcomings,this thesis mainly conducts the following research:(1)A boundary-assisted lightweight dual-resolution network is designed,called DRBANet,which can capture multi-scale semantic context and refine the result of semantic segmentation with the aid of boundary.At first,taking account of the high-level semantic and low-level detail feature maps are both important for semantic segmentation,a backbone is built with dual-resolution architecture to keep semantic features and detail cues.In addition,to extract multi-scale semantic features,a pyramid pooling module is designed.Finally,a boundary supervision is used in a lightweight manner,optimizing the result of semantic segmentation.(2)A lightweight two-resolution semantic segmentation network is designed,called DSNet,based on object boundary and semantic feature bilateral supervision.To solve the insufficient utilization of boundary features in DRBANet,a lightweight bilateral guidance module is designed,which realize the mutual guidance and constraint between boundary features and semantic one,meanwhile improve the prediction effect on small objects.Additionally,an improved pyramid pooling module is designed to integrate the multi-scale features step-by-step.Specifically,for the semantic feature maps capture form various-scale with different receptive fields,a hierarchical residual-like connections is used,to blend information in neighbor paths with lightweight convolution unit.(3)This thesis designs a boundary embedded lightweight dual resolution semantic segmentation network,called BENet.To relieve the problem of intra-class inconsistency and inter-class ambiguity,a boundary-aware criss-cross pooling operator is proposed.More specially,the predicted boundary contour divides semantic features into various regions with different classes,and then a strip pooling is conducted along vertical and horizontal direction within their own regions,making model obtain more robust semantic prediction around object edges and clear semantic labels inside object regions.