为了深入探究赤足足迹与人身性别之间的联系，本文以足迹的足型特征为基础，结合人工神经网络、足部生物力学及传统足迹学的相关理论，构建了一种经过遗传算法优化的、基于BP神经网络的赤足足迹人身性别分析模型。在本次实验中，我们共收集了500名年龄在20至30岁之间的受试者的赤足足迹图像，作为初始样本数据。这些图像经过Matlab图像处理软件的预处理，包括图像去噪、增强以及足迹特征的提取等基本操作后，被整理为实验数据集，并用于训练本文提出的分析模型。模型的训练结果表明，经遗传算法优化的BP神经网络分类，通过平面赤足足迹特征识别人身性别的准确率有很大提升，平均准确率达84.34%。此外，该模型还筛选出包含跖跟距、赤足跖宽等性别判定最具区分度的七大特征，为刑事技术人员在依据足迹全长、掌宽特征判断嫌疑人性别时提供了新的特征判断视角。实验得出的各项数据证明将人工神经网络用于足迹个体身份信息识别是基本可行的，通过遗传算法筛选的最优特征子集，显著提升了模型分类器的识别能力，其性能得以优化，这充分证明遗传算法优化神经网络在足迹分析中是有价值的，性别分类所对应的最优特征子集对足迹研究具有重要的学术和实践意义。

To explore the correlation between barefoot footprints and human gender, this study proposes a gender analysis model based on a Back Propagation (BP) neural network optimized by genetic algorithms. The model integrates footprint morphology, biomechanical principles, and traditional footprint analysis theories. A total of 500 barefoot footprint images from healthy subjects aged 20-30 were collected. These images underwent preprocessing using MATLAB software, including denoising, enhancement, binarization, and edge detection, followed by the extraction of 54 features encompassing grayscale distribution, geometric characteristics, and morphological parameters. Genetic algorithms were employed to select the optimal feature subset, identifying seven key discriminative features such as heel-to-ball distance (d_FT6), barefoot ball width (d_S1S2), and barefoot length (d_OP).

The optimized BP neural network classifier achieved an average accuracy of 84.34% in gender classification, significantly outperforming traditional BP networks (70.10%), logistic regression (75.62%), and support vector machines (78.94%). The analysis revealed that features like heel-to-ball distance and barefoot ball width exhibited strong sexual dimorphism, aligning with anatomical differences between genders. For instance, males generally displayed larger heel-to-ball distances and wider ball regions compared to females. Additionally, geometric features such as the rectangularity and aspect ratio of the metatarsal-arch-heel region further contributed to classification accuracy, reflecting structural variations in footprint morphology.

Statistical significance tests (e.g., t-tests and Mann-Whitney U tests) confirmed that the selected features demonstrated significant gender-related differences (p≤0.05), with high effect sizes (Cohen’s d>0.5) for critical parameters. Notably, while some statistical features (e.g., probability density functions) lacked individual significance, their nonlinear interactions within the neural network enhanced overall model performance. This highlights the importance of feature synergy in machine learning-based classification.

The study underscores the feasibility of applying artificial neural networks to footprint analysis, particularly in forensic contexts. The genetic algorithm’s ability to optimize feature selection and network parameters improved model robustness and generalization, addressing limitations of traditional methods that rely on linear assumptions or manual feature engineering. However, challenges remain, including sample homogeneity and potential errors in ink-based footprint acquisition. Future work should expand datasets to include diverse age groups, refine feature extraction techniques (e.g., incorporating pressure distribution data), and develop integrated software tools for real-world forensic applications.

In conclusion, this research advances the automation and objectivity of footprint-based gender analysis, offering both academic insights and practical value for criminal investigations. By bridging biomechanics, computer vision, and evolutionary computation, the proposed framework demonstrates the potential of machine learning in decoding complex biological patterns embedded in human footprints.