Machine Learning Based Specularity Detection Techniques To Enhance Indoor Navigation

This research investigated machine learning (ML) methods that detect glass specularity in direct and two-step approaches. We used YOLOv4, YOLOv4-tiny, and DeepLabv3 to detect glass, then developed a specular detection technique to detect specularity in images. We trained image segmentation and classification models to detect glass and fed the output as a masked image that only showed the detected areas from the previous step into our specularity highlight detection. Additionally, we developed the YOLOv4 and YOLOv4-Tiny models to directly detect specularity and evaluate the outcomes. We retrained our system by adjusting various parameters, such as image size and learning rate, to find the finest performance. The intersection over union (IOU) for the direct method of detecting specularity using YOLOv4 and YOLOv4-Tiny was 69.8% and 65.8%, respectively. The overall highest IOU results for specular detection in the two-step approach using YOLOv4 and YOLOv4-tiny architecture were 70.01% and 69.7%, respectively. The best overall result was produced by the two-step DeepLabv3-Specularity pipeline, with an IOU of 72.3%.