YOLOOI: YOLO with Robustness to Outliers and Illumination

Aluminum is a crucial component in modern society, widely used in fields such as transportation and manufacturing. It is characterized by its light weight, high strength, electrical conductivity, and corrosion resistance, making accurate assessment of its production quality highly significant. In industry, accurately identifying surface defects has become a mainstream task. Currently, defect detection on aluminum surfaces is generally based on image analysis. This strategy is cost-effective compared to other approaches and can leverage well-established technical methodologies. However, in real-world applications, variations in imaging results caused by different lighting conditions, along with unpredictable specular highlights, lead to a decline in detection performance. These challenges, combined with the interference from anomalous points, prevent existing methods from meeting practical needs. To address these issues, we propose the YOLOOI network, which incorporates an original Texture Emphasize and Supplement module (TES) for cross-scale feature fusion and a novel lighting-decoupling data augmentation strategy to eliminate the misleading effects of outlier spots. Our YOLOOI effectively handles variations in lighting conditions and mitigates interference from anomalous points. Experimental results demonstrate that our model significantly outperforms existing models on the aluminum surface defect dataset, highlighting the value of our work and its potential for further development in the future.