Nv-Net: Efficient infrared image segmentation with convolutional neural networks in the low illumination environment

Abstract

Infrared image segmentation is an essential problem in computer vision, especially when images suffer from the weak illumination and low-resolution problems. In this paper, we propose a novel fully convolutional neural network model for the low-resolution infrared images in weak illumination natural environments. First, the Nv-Net network is designed to segment the infrared images by introducing an enforcement layer in the front end of the framework. Then, a weighted-sigmoid-cross-entropy loss function is introduced to calculate the error between the prediction of the network and the ground-truth. To accelerate the network convergence, mean-variance normalization preprocessing is adopted. Moreover, a low illumination image dataset (LII) is built to train and test our model. The robustness and effectiveness of the proposed Nv-Net method are examined on the low illumination images with the mixed noises. Experimental results demonstrate that the proposed method has the flexibility to segment the arbitrary input images on several public datasets, such as the LII dataset, PASCAL VOC, and ADE20K. Compared with other state-of-the-art methods, the proposed Nv-Net method achieves the best segmentation performance in the low illumination environment.