Automated Transparent Glass Detection System using Semantic Segmentation

This paper puts forward a real time system for detecting transparent objects in live video feeds. Transparent object detection have unique challenges in computer vision due to absence of well-defines and structured visual features. In this paper we have tried to approach real-time detection using DeepLabV3+ model architecture along with edge-based detection techniques. The current model was developed to enhance distance calculation methods by utilizing features of segmentation maps and edge detection. That is use of semantic segmentation for each pixel classification of transparent glass and using numerous methods of image processing techniques. Current outcomes shows promise for real time application. Future work will focus on optimizing the model with a dataset of 5000 images.