Research And Application Of Transparent Object Semantic Segmentation Based On Fusion Of Edge Detection Networks

Semantic segmentation,as the cornerstone of image understanding,is one of the core tasks of computer vision.Although a lot of progress has been made in semantic segmentation technology based on deep learning,there are still some problems in some special scenarios or for segmentation of some special objects.For example,conventional segmentation methods are not effective or even difficult for transparent object segmentation tasks.To solve this problem,this thesis conducts an in-depth study of the semantic segmentation task of transparent objects in real scenes based on deep learning methods as follows:(1)Due to the special nature of transparent objects’ materials,the surfaces of their interior regions do not have fixed texture information and features,and their highly transparent properties make them often not easily observable.On the contrary,the edges of transparent objects have high contrast,and their boundary features are usually more obvious than the internal features.However,conventional semantic segmentation methods are highly prone to lose edge detail information that is crucial for transparent objects in the downsampling process.In this thesis,we propose a semantic segmentation network incorporating the edge detection module and the semantic segmentation module together for multi-task learning training,and use the edge features extracted by the edge detection module to compensate for the edge detail information lost during the downsampling process.Experiments on publicly available transparent object datasets show that the semantic segmentation network with the fused edge detection module can better improve the problem that conventional semantic segmentation methods do not work well on the transparent object task.(2)The highly transmissive nature of transparent objects can cause their subject regions to bring out the background parts behind them.With the interference of background imaging,this confusion of internal regions largely affects the consistency of the features extracted by the neural network,making it difficult for the neural network to accurately identify the features and semantic information extracted from transparent objects.For this reason,this thesis introduces an attention unit for the feature extraction module that can weaken the background noise interference and significantly enhance the foreground target features based on the fused edge detection module in Chapter 3.Experiments on publicly available transparent object datasets show that the method proposed in this thesis can further improve the intra-class consistency of the prediction results.(3)The semantic segmentation model of transparent objects proposed in this thesis is applied to a laboratory scenario,and a prototype system of transparent utensil segmentation in laboratory is designed and implemented.The system sample test results show that the prototype system designed in this thesis can meet the basic realistic requirements in a practical environment and has some practical application significance.