基于深度学习的重叠红外刑侦目标提取算法研究

Research on Overlapping Infrared Forensic Target Extraction Algorithms Based on Deep Learning

刑侦图像处理中,由于人活动的随意性和主观性,导致红外图像中存在手印重叠的问题,使得处理此类图像中重叠目标的提取较为复杂。针对重叠红外刑侦目标提取困难的问题,提出一个ResEU-Net网络模型,用于定位和分割手印目标区域以及重叠目标区域。此外,通过基于坐标变换和图像融合算法,有效地提取和复原单个重叠手印。通过算法实验结果及所提出网络模型的对比试验,验证了方法在提取红外重叠刑侦目标时的准确性和优越性。ResEU-Net网络模型的评估指标,包括mIoU、mAP以及准确率等,均优于其他模型。与其他目标提取方法相比,方法在提取整体手印目标、手印的重叠目标和单独手印目标时都具有较高的准确性和完整性。

In forensic image processing,the arbitrariness and subjectivity of human activities lead to the problem of fingerprint overlap in infrared images,making the extraction of overlapping targets in such images more complex.This paper addresses the difficulty of extracting overlapping infrared forensic targets and proposes a ResEU-Net network model for locating and segmenting fingerprint target areas as well as overlapping target areas.Furthermore,through coordinate transformation and image fusion algorithms,individual overlapping fingerprints are effectively extracted and restored.The algorithmic experimental results and comparative experiments with the proposed network model validate the accuracy and superiority of our approach in extracting overlapping infrared forensic targets.The evaluation metrics of the ResEU-Net network model,including mIoU,mAP,and accuracy,outperform other models.Compared with other target extraction methods,our approach demonstrates higher accuracy and integrity in extracting overall fingerprint targets,overlapping fingerprint targets,and individual fingerprint targets.