Content area
In power-system unstructured-data management, a large volume of images from inspection drones, substation cameras, and smart meters is heavily compressed due to bandwidth and storage constraints, resulting in lower resolution that hinders defect detection and maintenance decisions. Although deep-learning super-resolution (SR) techniques have made significant advances, real-world deployments still require a balance between reconstruction accuracy and model lightweightness. To meet this need, we introduce a channel-attention-embedded Transformer SR method (CAET). The approach adaptively injects channel attention into both the Transformer’s global features and the convolutional local features, harnessing their complementary strengths while dynamically enhancing critical information. Tested on five public datasets and compared with six representative algorithms, CAET achieves the best or second-best performance across all upscaling factors; at 4× enlargement, it outperforms the advanced SwinIR method by 0.09 dB in PSNR on Urban100 and by 0.30 dB on Manga109, with noticeably improved visual quality. Experiments demonstrate that CAET delivers high-precision, low-latency restoration of compressed images for the power sector while keeping model complexity low.