Infrared image technology is applied more and more widely in the field of criminal investigation. Heat traces left at the crime scene can be used to gather more evidence for the case. However, there are always some problems such as unclear contour and fuzzy extraction effect existing in infrared images. An image object extraction method based on improved U-Net to solve these problems is proposed in this paper. Three layers of up-sampling and down-sampling are adopted in the network structure. Double-cubic interpolation is used for up-sampling, and 3×3 convolution with step size of 2 is used for down-sampling. The shallow feature and deep semantic information are combined by skip connection. The introduction of Dropout and Batch Normalization structures makes the network converge faster and better. Taking infrared fingerprint of criminal investigation as the research object, the image object extraction effects of U-Net, Sobel operator, watershed, maximum entropy and Otsu are improved by comparing experiments. The results show that the U-Net network built in this paper can extract the contour information of infrared fingermarks more completely and effectively, and good results are achieved in the extraction of infrared fingermarks.