Jar-RetinexNet: A System for Non-Uniform Low-Light Enhancement of Hot Water Heater Tank Inner Walls

The manual inspection of electric water heater enamel is inefficient and unreliable, a challenge stemming from the tank's narrow (approx. 50 mm) aperture that creates extremely dim, non-uniform lighting. Existing enhancement algorithms struggle with such complex industrial imagery. To address this, we propose an integrated hardware-software system: the three-axis Image Acquisition Robot (IAR) and Interactive Visualization Enhancement Software (IVES). Using this system, we constructed and released the first Heater Tank Inner Wall (HTIW) dataset, containing 900 real-world images. We further introduce jar-RetinexNet, a Retinex-based network featuring a Feature Preservation Attention Module (FPAM), a Cascaded Channel-Spatial Attention Module (CSAM) for precise decomposition, and a Random Affine Generation (RAG) module for generalization. Experiments show that jar-RetinexNet significantly outperforms state-of-the-art methods, achieving the best no-reference quantitative scores on our HTIW dataset: a BRISQUE of 25.4457 and a CLIPIQA of 0.3160.