Detection of cervical vertebrae from infrared thermal imaging based on improved Yolo v3

Yolo has achieved great success in the field of image segmentation, and has been applied to infrared thermal imaging detection. However, in the feature pyramid for feature fusion, high-level spatial feature information is lost, and both high-level and low-level features have poor semantics. This paper proposes an infrared thermal imaging cervical spine part extraction method based on improved Yolo v3. In order to make up for the channel information lost in feature fusion, this paper convolves the high-level features, and then enhances the residual features to reduce the semantic loss caused by the number of channels by compensating for the spatial context information. To reduce the semantic gap of additive fusion, this paper applies an attention mechanism on low-level features. The improved Yolo v3 algorithm was used to extract the cervical vertebrae in infrared thermal images, and comparative experiments were completed. Experiments on the dataset collected in the cooperative hospital demonstrate that our proposed improved Yolo v3 achieves better performance.