Object Detection and Recognition for Visually Impaired Users: A Transfer Learning Approach

Vision is one of the most important human senses, and it plays a critical role in understanding the surrounding environment. However, millions of people in the world are experiencing visual impairment. They are facing difficulties in their daily navigations since they cannot see the obstacles in their surroundings. In this paper, we applied transfer learning to two pre-trained models, Single Shot Detector Mobilenet V1 (SSD) and Faster Region-Convolutional Neural Network Inception V2 (Faster R-CNN), to detect and recognise 40 classes of common objects in surroundings. The performance of these two trained models was subsequently compared. The fine-tuned Faster R-CNN model was able to achieve an mAP (Mean Average Precision) score of 0.8961 with an inference time of 7.2 seconds. On the other hand, the SSd model achieved an mAP score of 0.5708 with an inference time of 1.8 seconds. In consludion, Faster R-CNN appears as the suitable model to be implemented in the mobile application to help visually impaired people as it produces a good mAP result.