Channel Split Convolutional Neural Network (ChaSNet) for Thermal Image Super-Resolution
The ability of thermal sensors to penetrate smoke, mist, dust and aerosol makes them attractive for deployment in essential applications in military, medical, agricultural and animal ecology over the regular optical cameras. However, unlike optical imaging devices, the most sophisticated commercial thermal imaging sensor does not match the megapixel imaging ability. The Low-Resolution (LR) images from thermal sensors can be enhanced through a software-driven solution called Super-Resolution (SR). A number of works have been proposed to employ deep networks for SR task; however, they are overloaded with redundant features due to the deep architecture. This paper introduces a Channel Splitting-based Convolutional Neural Network (ChasNet) for thermal image SR eliminating the redundant features in the network. The use of channel splitting extracts the versatile features from Low-Resolution (LR) thermal image, helping to preserve high-frequency details in the SR images. We demonstrate the applicability proposed network for SR task in two different scenarios organized in the PBVS-2021 Thermal SR Challenge, consisting of noise elimination (Track-1) and domain shifting (Track-2). The efficacy is justified by comparing the SR results with other state-of-the-art thermal SR techniques in qualitative and quantitative metrics. A set of extensive experiments separately analyzes the importance of each block in the proposed architecture. The code of this work is also published online.