Abstract: Horizontal silos are crucial in China’s grain storage system. Surface defects directly compromise structural integrity and storage safety. Traditional manual inspection of horizontal silo surface suffers from low efficiency and frequent missed detection, while existing recognition methods lack sufficient robustness in scenarios with or without surface coatings and complex backgrounds. This study develops an intelligent surface defect identification method for horizontal silos and develops its engineered application system. Based on YOLOv5s model, C3-SE-SimAM and CBAM-Spatial attention modules are introduced into the backbone and neck structure, respectively, to enhance the extraction capability of weak defect features and suppress complex background interference. A MobileNetV3-small-based pre-classifier is designed to perceive the surface coating state and drive adaptive routing of the detection branch, which alleviates the feature distribution shift caused by coating differences. Meanwhile, a detection system supporting multi-source input and automatic report generation is developed based on PyQt5. A dedicated dataset (GWS-DD) containing 2 740 images for coating classification and 2 740 images for defect annotation was constructed. On the coated surface test set (D1), the improved model achieved a mean average precision (mAP) of 0.812, a recall of 0.827, a precision of 0.835, and an F1-score of 0.831. On the uncoated set (D2), the model achieved mAP, recall, precision, and F1-score values of 0.808, 0.812, 0.814, and 0.813, respectively, representing a 24.9% mAP improvement over the baseline YOLOv5s model across the two test sets. Comparative experiments showed that the proposed method outperforms Faster R-CNN, YOLOv7-tiny, and YOLOv8s, with mAP gains ranging from 25.7% to 35.6%. Cross-domain validation revealed that the model trained on coated surfaces suffered at 36.8% mAP drop when directly applied to uncoated surfaces, while the reverse application resulted in a 41.2% mAP decline. In contrast, the proposed coating-aware adaptive routing improved the mAP by 12.5% and 11.9% on coated and uncoated surfaces, respectively, compared with the unified model trained on mixed datasets. The integrated system achieved a frame rate of over 60 FPS on an RTX 5060 Ti GPU, reducing the manual silo inspection time from 2-3 hours to less than 10 minutes. By synergistically integrating lightweight attention mechanisms and a coating-aware adaptive pipeline, the proposed method effectively alleviates feature distribution shifts. The developed system greatly improves inspection efficiency and standardization, providing solid technical support for the daily maintenance and safe operation of grain storage infrastructure.