Thermal infrared image semantic segmentation for night-time driving scenes based on deep learning

Semantic segmentation of thermal infrared (ThIR) images is challenging because the images considered in this task are highly complex. The discrimination of image regions is very difficult, and the traditional techniques fail to discover the crucial semantic information from the images completely. To overcome such issue, this paper introduces a novel network model for ThIR image semantic segmentation that facilitates effective image-to-image translation and reduces semantic encoding ambiguity. The proposed model is named top-down attention and gradient alignment-based graph neural network (AGAGNN). A top-down guided attention module (GAM) is utilized in the proposed model to deal with semantic encoding ambiguity. Apart from this, an elaborate attention loss is introduced to ensure a hierarchical coding of features. Also, the edge distortion problem due to the translation of images is reduced with an organized gradient alignment loss. The proposed model is evaluated under the Python platform based on pixel-level annotations over the KAIST dataset. The proposed model has shown 98.3% accuracy, and the comparative analysis has proved that the model is more effective than the existing models in preserving semantic information.