A Three-Stage Low-Illumination Image Enhancement Method Based on Feature Refining and Its Application in Inspection Robot for High-Voltage Substation Room

Bo Fu, Hongguang Wang, Yong Jiang, Yifeng Song

Abstract: To address color deviation and noise problem in some Retinex theory based image enhancement methods, this paper proposes a three-stage low-illumination image enhancement method based on feature refining. Additionally, a vision subsystem for an electric inspection robot is introduced based on this network to mitigate the impact of low illumination environments on vision detection tasks in high-voltage substation rooms. In the first stage, the proposed network decomposes the image using Retinex theory and leverages the Swin Transformer to perceive global information. In the second stage, the image undergoes preliminary enhancement. In the third stage, the feature refining network is introduced. Initially, a U-shaped network is employed to extract image features. Subsequently, a convolutional neural network and attention mechanism are applied to screen the extracted information, completing the feature refinement and improving the quality of the result. Through illumination assessment and algorithm selection, the proposed vision subsystem for the robot effectively addresses the challenges of low illumination enhancement and object detection in complex lighting conditions in high-voltage transformer rooms. Experimental results demonstrate that the proposed network outperforms some state-of-the-art methods in terms of peak signal-to-noise ratio and structural similarity index. Moreover, the vision subsystem presented in this paper significantly enhances the robot's detection capabilities, reducing the omission ratio and fallout ratio, thus validating the effectiveness of the proposed method.