Litter Detection from Digital Images Using Deep Learning

In order to achieve automatically litter detection in residential area, machine vision has been applied to monitor environment of surveillance. Based on our observations and comparative analysis of the current algorithms, we propose an improved object detection method based on Faster R-CNN algorithm and achieve more than 98% accuracy of litter detection in surveillance. Through our observations, most of litters are small objects, we apply feature pyramid network to Faster R-CNN and optimize it by merging different layers by using multiply operate. Besides, we replace cross-entropy loss function with focal loss function to solve the problem of anchor imbalance by using region proposal network (RPN) and offer attention module through RPN to feedback the whole network. We collected more than 8000 labeled images from our surveillance videos for model training. Our experiments show that the improved Faster R-CNN achieves a satisfied performance in real scene. © 2022, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.