Abstract
Unmanned aerial vehicle (UAV) inspection provides a flexible means of monitoring geographically dispersed equipment in distribution networks; however, the limited resolution, motion-induced blur, transmission constraints, and high cost of precision infrared sensors can degrade image perception and weaken subsequent inspection decisions. To address these problems, this paper proposes an adaptive super-resolution UAV inspection and image perception strategy for distribution networks. A compressed-sensing-based non-blind super-resolution model is developed to recover high-resolution infrared information from degraded images acquired by low-cost UAV-mounted sensors. The proposed method first employs a Gaussian prior to obtain a stable preliminary deconvolution result and then generates a structural label image through threshold shrinkage. Based on the differences between salient equipment contours, thermal-transition boundaries, and smooth background regions, a hyper-Laplacian prior is adaptively assigned to preserve fault-related details while suppressing noise, ringing, and reconstruction artifacts. A dual-prior quadratic estimation mechanism is further introduced to automatically regulate the penalty coefficients and improve reconstruction stability under changing degradation conditions. The reconstructed images are subsequently used for infrared-visible feature matching and equipment perception, providing clearer visual evidence for identifying suspected thermal anomalies, prioritizing inspection targets, and planning UAV reinspection strategies in distribution networks. Experiments involving 60 synthetic reconstruction cases and practically captured infrared images show that the complete method achieves an average PSNR of 32.41 dB, an SSIM of 0.914, an average gradient of 7.83, and an information entropy of 7.21. Its mean correct feature-point matching rate reaches 80.72%, demonstrating improved preservation of equipment-discriminative structures. Moreover, across 165 mismatched blur kernels, the maximum performance loss is approximately 3.6%, confirming its robustness to moderate kernel-estimation errors. These results indicate that the proposed strategy can enhance UAV infrared image quality and provide a reliable perception basis for adaptive inspection planning, abnormal-equipment localization, and condition-oriented operation and maintenance of distribution networks.