Improving night time road safety deep learning for semantic segmentation and object detection in thermal infrared images

Abstract

This research addresses two critical aspects of image processing within the domain newlineon object detection as well as semantic segmentation at both thermal infrared and urban newlinelandscape images, aiming to improve the accuracy, efficiency, and robustness related to newlinesystems for recognizing images. newlineThe first part of this study focuses on the challenging task of semantic segmentation in thermal infrared images, which are complex and difficult to analyze due to the newlineintricate nature of their features and the ambiguity in semantic encoding. Traditional newlinetechniques fall short in fully capturing the crucial information within these images. newlineTo overcome this, we propose a novel network model named the top-down attention newlineand gradient alignment-based graph neural network, designed specifically for thermal newlineinfrared image semantic segmentation. The top-down attention and gradient alignmentbased graph neural network model integrates a top-down guided attention module to newlinetackle semantic encoding ambiguity effectively and introduces an attention loss function to protect feature coding within a hierarchical manner. Additionally, a gradient newlinealignment loss is incorporated to address the edge distortion problem during image newlinetranslation, significantly improving the quality of the segmented outputs. The model is newlineevaluated using pixel-level annotations from the KAIST dataset, achieving an impressive accuracy of 98.3%. Comparative analysis demonstrates that the top-down attention newlineand gradient alignment-based graph neural network outperforms existing ones for maintaining semantic data and provides more reliable and accurate thermal infrared imaging newlinesegmentation findings. newlineIn the second part of the research, we turn our attention to the problem of precise newlineitem identification in urban environments, particularly in nighttime infrared images. newlineEnhancing hue in thermal images is a common focus of present-day models, but they newlinefail to effectively detect objects in complex urban settings. To narrow this disparity, newlinewe suggest an innovative strategy that uti