3H-ObjectDet: A Target Detection Method Based on Haar Wavelet Transform Downsampling and Hybrid Attention Mechanism for Medical Fracture Image

In recent years, in order to reduce the incidence of misdiagnosis and missed diagnosis in fracture diagnosis and maximize the protection of patients' lives and health, deep learning has achieved rapid development in the field of fracture diagnosis. Target detection models can obtain the location and category of fractures in images simultaneously, and the detection results output are more intuitive, making it convenient for clinical doctors to use directly. However, the following difficulties have troubled clinical use: (1) inefficient extraction of image features, (2) difficulty in accurate detection classification, and (3) large calculation volume causing delayed detection output. This paper proposed a medical fracture image target detection method based on wavelet transform downsampling and mixed attention mechanism, and made relevant innovations: (1) introduced a mixed attention mechanism instead of traditional convolution operation in the feature extraction of the backbone network, (2) innovated the detection classification and localization module to effectively identify features, and (3) simultaneously enhanced feature extraction ability through wavelet transform downsampling and reduced computation. The innovative detection model achieved a 21.2% increase in accuracy and a 10.1% increase in mAP@0.5 in the rib fracture dataset, while reducing computation by 14.8%. In the upper limb fracture and the hand fracture dataset, the high performance indicators have also been significantly increased, showing the high efficiency and generalization of the innovative detection model.