The aim of this study is to solve the problems of blurred edges and incomplete contours in infrared image criminal investigation scenes. A method of extracting fuzzy criminal investigation targets based on feature enhancement is presented in this paper. Firstly, the BCES-Net network model is designed and the feature images with strong semantic information are obtained by using STCAM. Then the edge features and fuzzy criminal investigation target features containing semantic category information are obtained by modeling extraction. In the training process, based on specific loss functions and multiple feature fusion techniques, the segmentation performance of edge and fuzzy criminal investigation targets is improved through repeated supervised learning and training correction. In the hand heat trace data set, compared with DeeplabV3+, U-Net, HRNet, PSPNet and other models, BCES-Net is significantly superior in mIoU, mAP, accuracy and other evaluation indexes. mIoU reaches 88.3%, mAP reaches 94.35%, and the accuracy reaches 95.5%. This research innovatively improves the extraction accuracy of fuzzy infrared criminal detection targets and provides technical support for practical application.