Hierarchical detection of unsafe conditions in banks based on YOLOv8

This study proposes an improved YOLOv8 algorithm based on the network model framework in response to issues in grading accuracy, slow speed, high false alarm rate, and the excessive workload of monitoring staff encountered in the tiered monitoring of unsafe banking conditions. These challenges lead to less effective monitoring results. The study utilized a dataset that includes PASCAL VOC2007 and a section of a self-built dataset, collectively comprising of 8,511 images. Initially, a Partial Convolutions module was introduced to the backbone section, enhancing the ability to recognize occluded objects in images and improving the resistance to interference. Following that, an enhancement was made to the Bottleneck structure, integrating Large Kernel Attention (LSKA) and enhancing its detection capabilities for small targets. For large target detection, Multi-Scale Dilated attention was employed to enhance the model's processing efficiency and detection accuracy. Lastly, an ASFF head was incorporated. Overall, the improvements resulted in a 3.0% increase in mAP@0.5, boosting detection accuracy and speeding up the detection process.