Research On Two Stage Detection Method Of Transparent Objects In Complex Environments

Transparent object detection aims at detecting transparent objects such as glass cups and glass doors from images,and it is widely used in smart home,unmanned logistics and other fields.However,due to the particularity of the material of transparent objects,their imaging is particularly susceptible to environmental factors such as background and illumination,which leads to the fact that existing object detection algorithms are prone to missed detections and false detections when detecting transparent objects in complex environments.For a transparent object,the perception process of human vision can be divided into the following two stages: 1)Combine the overall information of the object to be detected and the surrounding local detail information to comprehensively predict whether it is a transparent object;2)Predict the specific category of the transparent object according to the its features.Based on the above-mentioned human visual perception mechanism,this paper designs a two-stage detection algorithm to achieve high-performance detection of transparent objects in complex environment.The main contributions of this paper are as follows:(1)This paper Analyze the factors that affect the performance of the transparent object detection model,and establish our own transparent object detection dataset based on this.(2)This paper first designs a master-slave bipartite network structure to enhance the feature extraction capability of the backbone.Then,based on the first stage of human visual perception transparent object,a feature pyramid structure based on fine-grained feature enhancement and receptive field enhancement is designed,and combined with the self-attention mechanism,the features extracted by the backbone network are further enhanced and fused,so as to expand the field of view and enhance the object details.(3)Based on the second stage of human visual perception,this paper designs a dual-branch detection network structure based on deformable region of interest pooling,which uses the pooling of deformable regions of interest to improve the object feature representation ability,and combines bounding box regression and object recognition dual-branch network structure to improve the bounding box accuracy of the transparent objects in the case of deformation,occlusion or overlap.In order to verify the effectiveness of the proposed algorithm,this paper conducts experiments based on our own transparent object detection dataset.The experimental results show that compared with the baseline model,our proposed two-stage transparent object detection method achieves a 3.6% AP improvement.Compared with other mainstream object detection algorithms,our method also shows better performance.Figure 34,table 8,references 99.