Abstract To address these challenges, this paper proposes a power equipment image reconstruction network that integrates adaptive convolution and a lightweight transformer (LT). An adaptive local feature extraction module based on convolution is designed, which leverages dynamic weighting to fuse fine-grained local features with varying receptive fields from both high- and low-frequency image components, thereby enhancing the reconstruction of local textures. An LT architecture tailored for power image reconstruction tasks is developed, in which the traditional global self-attention mechanism is decomposed into a locally aware attention guided by global information. This approach effectively reduces computational complexity while preserving global perceptual capability.
Zha et al. (Thu,) studied this question.