R-Net: A novel fully convolutional network–based infrared image segmentation method for intelligent human behavior analysis

Abstract

Infrared weak-illumination image segmentation is a complex task in computer vision and intelligence education. In this work, we proposed a novel convolutional neural network for infrared image segmentation, which can overcome the problems of motion blur, low resolution, and random noise. In particular, we proposed a new loss function that considers the shape, area, and centroid during learning and integrates them into a simple deep learning model. We evaluated our method on our annotated dataset (Night Human in the Wild Scene dataset), comprising approximately 22,338 natural low-quality images corrupted by mixed noises, and several public datasets with high-resolution images, such as ADE20K and PASCAL VOC. Experimental results show that the proposed R-Net outperforms the existing infrared image segmentation models.