【摘要】 针对刑侦现场标尺物证图像因拍摄角度、镜头畸变导致的几何失真问题，提出一种融合深度学习与亚像素定位的畸变校正方法。利用YOLOv5神经网络自动分割标尺图像中的圆形和矩形特征；采用Hough变换和连通域分析精确定位特征形心，在此基础上引入基于二次插值的亚像素定位技术提升选点精度；后通过透视变换实现图像等大校正。实验结果表明，该方法在2万枚物证图像数据集上实现了30 cm量程内误差小于5 mm的校正精度，兼具深度学习对环境变化的强鲁棒性和亚像素定位的高精度优势。更多还原

【Abstract】 To address the geometric distortion in scale evidence images caused by shooting angles and lens distortions in crime scene investigations, a distortion correction method integrating deep learning and subpixel localization is proposed.First, the YOLOv5 neural network is used to automatically segment circular and rectangular features in the scale image.Then, Hough transform and connected component analysis are applied to accurately locate the centroids of these features.On this basis, a subpixel localization technique based on quadratic interpolation is introduced to enhance point selection accuracy.Finally, perspective transformation is employed to achieve uniform image correction.Experimental results demonstrate that the proposed method achieves a correction accuracy of less than 5 mm within a 30 cm range on a dataset of 20 000 evidence images, combining the robustness of deep learning to environmental variations with the high precision of subpixel localization.更多还原