Design and Experiment of a Glass Detection System Based on Context and Edge Attention Using Deep Learning Methods

The reflective and refractive properties of transparent glass, combined with its interactions with other objects, introduce challenges for the semantic segmentation of transparent glass scenes. Although existing methods utilize Fully Convolutional Networks (FCN) for segmentation, these models possess large parameter sizes, leading to low feature utilization, wasteful memory consumption, and intensive computational resources. Moreover, these models easy to lose the edge information specifically in transparent object scenes. In the paper, a glass detection system based on context and edge attention using deep learning methods are proposed. Firstly, the contextual information and edge attention are combined. The contextual information enhances the semantic understanding of transparent objects and the edge attention highlights the edge features of transparent objects. Secondly, the edge feature group module and edge attention module to effectively capture various boundary details and attain precise segmentation. Thirdly, the image context information and resolution are decreased by implementing the Atrous Spatial Pyramid Pooling (ASPP) and dense connectivity module, and the model complexity and parameter count are decreased. The experimental results validate the effectiveness of the method with reductions in both Mean Absolute Error (MAE) and Balanced Error Rate (BER) to 0.068 and 4.41%, respectively. Additionally, Pixel Accuracy (PA) and Intersection over Union (IOU) demonstrated improvements to 92.51% and 89.64%, respectively.