Multi-Scale Information Attention Network-Based Lightweight Super-Resolution Reconstruction For Infrared Images of Power Equipment

Abstract

This paper presents a super-resolution technique based on a multi-scale information attention network (MIA) aimed at tackling the existing issues of low resolution and unclear temperature distribution in infrared images of power equipment. Firstly, a multi-scale residual distillation module (MRD) is constructed by distilling and refining the input feature maps using lightweight blueprint separable convolution (LBSConv) to fully extract the multi-scale feature information of power equipment; secondly, mixed local-global channel attention (MLGCA) is built to adaptively capture the local and global information in the spatial dimensions of the power equipment, and to enhance the key features extracted by the model then, the super-resolution network is constructed so that the network focuses on multi-scale information, such as texture details and overall contours of infrared images of power equipment, to improve the feature expression ability of the model. Experimental results demonstrate that the proposed MIA achieves a peak signal-to-noise ratio (PSNR) of 35.315 dB and an structural similarity (SSIM) of 0.9543 on the power equipment infrared image dataset, which are better than several other lightweight models, and provide a better effect in reconstructing power equipment during subjective visual evaluation. The proposed method can enhance the thermal imaging resolution for power equipment, which supports power equipment fault diagnosis.