Super-Resolution Reconstruction of Motor Long-Wave Infrared Images Based on Improved USR-Net

Temperature analysis of long-wave infrared images serves as an effective approach for motor fault detection. However, current long-wave infrared cameras suffer from relatively low imaging resolution. To overcome this limitation, this paper proposes a super-resolution reconstruction method that emphasizes the restoration of complex edge structures in motor images. The method first applies adaptive Wiener filtering to eliminate environmental noise and subsequently incorporates a spatial attention mechanism (SA) within the residual blocks of the USR-Net to enhance edge feature extraction. An improved adaptive channel attention mechanism (Simple-Se) is introduced after the decoder to further refine the reconstruction of complex edges. The reconstruction performance is further optimized by combining perceptual loss and mean squared error loss. Experimental results indicate that, for 2x degraded images, the proposed method achieves a Peak Signal-to-Noise Ratio (PSNR) value exceeding 41 dB, outperforming other methods, with the Structural Similarity Index (SSIM) reaching 0.9872. Furthermore, the Learned Perceptual Image Patch Similarity (LPIPS) value for 2x degraded images is below 0.095. Overall, the proposed method demonstrates outstanding performance in image reconstruction quality, detail preservation, and visual effects, making it highly suitable for high-precision image restoration in practical applications requiring superior image quality.