Dense bone stick text detection algorithm in complex texture background

Bone stick is an important cultural relic that records the products handed over by local officials to the central government in the Western Han Dynasty. It is of great significance to accurately detect the written content. In order to solve the problem that bone stick text is difficult to extract under complex texture background and the dense text and adhesion lead to multiple characters in one frame, a bone stick text detection algorithm combining self-attention convolution and improved loss function is proposed. Firstly, a self-attention convolution module is added to the YOLOv5 feature extraction to enhance the network's attention to the features of bone stick, and to make the model capture more global information and suppress the interference of the crack to the feature extraction. In addition, the Focal-EIOU loss function is used to replace the CIOU network for optimization. Focal-EIOU uses the wide-height loss to reduce the wide-height gap between the prediction box and the real box, and eliminates the prediction box larger than the real box, the detection frame redundancy problem caused by text density and adhesion is solved to improve the precision prediction ability of the model. The experimental results show that the average accuracy of the proposed algorithm reaches 93.35%, which is 3.08% higher than that of YOLOv5. It is more suitable for the task of detecting dense adhesive bone stick text in complex texture background.