Novel deeper AWRDNet: adverse weather-affected night scene restorator cum detector net for accurate object detection

• Published in: Neural computing & applications Vol. 35; no. 17; pp. 12729 - 12750
• Main Authors: Singha, Anu, Bhowmik, Mrinal Kanti
• Format: Journal Article
• Language: English
• Published: London Springer London 01.06.2023; Springer Nature B.V
• Subjects: Artificial Intelligence; Artificial neural networks; Aspect ratio; Boxes; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer networks; Computer Science; Computer vision; Data Mining and Knowledge Discovery; Datasets; Feature maps; Fog; Frames (data processing); Image Processing and Computer Vision; Night; Object recognition; Original Article; Performance measurement; Probability and Statistics in Computer Science; Weather; Deeper layers; Convolutional block; Adverse weather; Adaptive max pooling; TU-VDN; Object detection
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-023-08390-7

Object detection in adversarial atmospheric attacks, such as fog, rain, low light, and dust conditions, is a challenging task with regards to computer vision. Moreover, the applicability of convolutional neural network-based object detection architectures in various weather-affected night-time thermal scenes has not been extensively reported in recent and past literatures. The extraction of region of interest through anchors from each multi-resolution feature map (FM), either shallow or deep, suffers from several issues in adverse weather-degraded scenarios. Our proposed architecture, namely adverse weather-affected night scene restorator cum detector net (AWRDNet), focuses on the process of recovering such adverse weather-degraded video frames to restored frames through deeper convolutional layers. Further, our network reduces the time-consuming generation of pre-defined anchors in each FM at a deeper de-convolution layer, which combines different scales and aspect ratios for anchor boxes from multiple sets to naturally handle objects of various sizes. Considering the multi-scale anchor boxes at multiple set, an anchor refinement strategy has been applied to reduce memory consumption. The performance of the AWRDNet architecture is evaluated using standard detection performance metrics over the Tripura University Video Dataset at Night Time (TU-VDN) dataset which contains objects with annotated bounding box of the image frame sequences, and the available PASCAL VOC 2007 2012 datasets, and ZUT thermal dataset.