This paper analyzes the importance of optimizing fingerprint images from the perspective of front-line forensics personnel. Due to defects such as fingerprint ridge blurring and fingerprint image overlap, it may cause delays in fingerprint identification and lead to misjudgments (false positives, false negatives) and other consequences. This paper simulated common fingerprint images (fuzzy feature points) at criminal cases, and captured, segmented, and reconstructed fingerprints. Two deep learning model architectures, U-net [1] and FCN [2], were applied to realize image segmentation and image reconstruction. The first task is fingerprint image segmentation, where the fused images are segmented into incomplete fingerprints. The second task is to reconstruct the complete fingerprint from the image obtained after segmenting the incomplete fingerprint. Field verification for fingerprint image reconstruction and separation on red envelope were carried out. Ethyl cyanoacrylate method is used to visualize paper latent fingerprints and would produce “white” fingerprints, which is not easily distinguished from the red background. The pre-processing of fingerprints on red paper background is the key to the accuracy of fingerprint recognition. U-net image processing technology demonstrates the contributions of the forensic personnel in processing fingerprint images are as follows: [1] Improve efficiency. [2] Reduce misjudgment.