Aiming at the challenges of low illumination (< 10 Lux) and complex weather in power inspection at night, which lead to the sharp drop of signal-to-noise ratio (SNR) of visible light imaging and the decline of single spectrum detection performance, a depth adaptive algorithm for anomaly detection in power inspection under full spectrum night vision is proposed. This algorithm designs an Adaptive Attention Module (AAM) that generates dynamic weights via cross-modal similarity matrices, effectively resolving the inconsistency issue in multispectral feature spaces. It employs Neural Architecture Search (NAS) to optimize a lightweight-high-precision dual-branch backbone network, maintaining feature extraction capabilities while compressing parameter counts. Dynamic preprocessing and frequency domain enhancement technology based on illumination perception are introduced to improve the SNR˗ of low-quality images, Through the dynamic reasoning mechanism driven by risk assessment, the high-risk target is automatically switched to the high-precision branch to achieve the balance between speed and accuracy. The PSN-Inspection data set containing 15,200 full-spectrum image triplets is constructed for verification. The experimental results show that the algorithm email protected reaches 84.7%, and the small target detection email protected reaches 73.8%, which is 13.2 percentage points higher than the traditional mosaic method, with an average processing speed of 52 FPS, and shows stronger robustness in bad weather such as fog, rain and snow.
Yan et al. (Sun,) studied this question.