SH-Detection: An Efficient YOLO-Based Framework for High-Accuracy Lockhole Detection

Accurate detection of container lockholes is crucial for precise container positioning in automated systems. However, the small size of the lockholes and the complex background present significant challenges for standard detection models, especially when using conventional downsampling methods that risk losing key features. This paper introduces an efficient and accurate lockhole detection method, which enhances the YOLO model and employs a sub-image overlapping fusion algorithm. The method is divided into two main stages: model improvement and image fusion. In the first stage, we enhance the YOLO network through a multi-level feature representation mechanism, propose a novel IoU calculation method to prevent gradient vanishing during training, and introduce an attention mechanism to focus on key corner regions of the container. In the image fusion stage, we mitigate the loss of detail caused by image downscaling by dividing the original image into overlapping sub-images, ensuring comprehensive detection. The overlapping areas are carefully managed with a custom algorithm to avoid duplicate detections. This method significantly improves the accuracy of lockhole detection while maintaining computational efficiency, outperforming existing approaches in both robustness and precision.