【摘要】 犯罪现场遗留血迹的提取与鉴定为案件侦破提供了重要依据，但其快速、无损显现与检验仍是法庭科学领域的研究热点。为提高血迹的显现效率与检测精度，高光谱技术逐渐被应用于血迹的无损识别。然而，现有高光谱成像技术在血迹及类血迹识别中，尤其是针对复杂客体上的血迹时，存在识别准确率低、效率不足等问题。为此，通过将SENet通道注意力机制与一维残差网络（ResNet18-1D）融合，提出了一种基于高光谱成像技术的血迹识别模型，旨在提高血迹的高光谱成像技术识别准确率和效率。SENet通道注意力机制通过学习自动获取每个特征通道的重要性权重，进而增强有效特征并抑制无关特征。针对承痕客体的复杂性，改进传统SENet模块，采用双分支瓶颈模型以提升模型的适用性。为了应对公安实战中复杂多变的情况，在包含多种承痕客体的公开数据集Blood Detection数据集上进行了两组实验。（1）高光谱直推式分类实验场景：训练集与测试集来自同一高光谱图像，重点分析承痕客体对血迹特征的干扰。实验表明，该模型在包含复杂客体的模拟场景中总体精度（OA）与平均精度（AA）分别达96.8%和97.6%,较当前最优的Hybrid CNN模型提升1.3%和1.9%;（2）高光谱归纳式分类实验场景：在基础场景训练后直接迁移到另一幅图像中测试，此实验侧重血迹及类血迹物质的预识别能力，更具挑战性，但更符合实际应用场景。实验表明，模型总体精度与平均精度分别为63.3%和65%,较当前最优的RNN模型提升2.2%和1.6%,并通过错误源分析发现，番茄汁因其在470 nm附近的吸收峰与血液在415 nm的特征吸收峰相似，成为主要干扰源。除横向对比不同算法外，还通过消融实验验证了SENet通道注意力机制模块对模型性能的影响。结果显示，改进后的SENet通道注意力机制模块相较于原始SENet通道注意力机制模块，使模型在两种分类场景中的总体精度与平均精度都有提升。同时，效率测试显示虽然该模型参数量较大，但残差结构与双分支SENet的协同设计显著降低计算成本，训练耗时仅为45 ms·epoch^-1,满足实战的效率需求。更多还原

【Abstract】 The extraction and identification of bloodstains left at crime scenes provide an important basis for case investigation, but their rapid and non-destructive development and examination remain a research hotspot in the field of forensic science. To enhance the development efficiency and detection accuracy of bloodstains, hyperspectral technology has gradually been applied to the non-destructive identification of bloodstains. However, existing hyperspectral imaging techniques have limitations, including low recognition accuracy and insufficient efficiency in identifying bloodstains and blood-like substances, particularly when dealing with bloodstains on complex objects. Therefore, a bloodstain identification model based on hyperspectral imaging technology by integrating the SENet channel attention mechanism with a one-dimensional residual network（ResNet18-1D） was proposed in this paper, aiming to improve the accuracy and efficiency of bloodstain recognition by hyperspectral imaging technology. The SENet channel attention mechanism automatically acquired the importance weight of each feature channel through learning, thereby enhancing effective features and suppressing irrelevant ones. In view of the complexity of the trace-bearing object, this paper improved the traditional SENet module and adopts a dual-branch bottleneck module to enhance the applicability of the model. To address the complex and dynamic nature of forensic practice, this paper conducted two sets of experiments on the public blood detection dataset, which contains multiple traceable objects.（1） Hyperspectral Transductive Classification Scenario. Both training and test sets were derived from the same HSI image. This experiment focused on analyzing substrate interference with bloodstain spectral features. Results show the model achieved an overall accuracy（OA） of 96.8% and an average accuracy（AA） of 97.6% in the complex simulated scenario, representing improvements of 1.3% and 1.9%, respectively, compared to the state-of-the-art Hybrid CNN model.（2） Hyperspectral Inductive Classification Scenario. The model trained on the baseline scenario was directly transferred to test on a different image. This experiment focused on the pre-identification capability for bloodstains and blood-like substances, presenting greater challenges but better reflecting real-world application needs. Experiments showed that the overall accuracy and average accuracy of the model were 63.3% and 65% respectively, which were 2.2% and 1.6% higher than the current optimal RNN model. Through error source analysis, it was found that tomato juice, due to its absorption peak near 470 nm being similar to the characteristic absorption peak of blood at 415 nm, has become the main interference source. In addition to making horizontal comparisons of different algorithms, this paper also verified the impact of the SENet channel attention mechanism module on model performance through ablation experiments. The results showed that the improved SENet channel attention mechanism module, compared with the original SENet channel attention mechanism module, had enhanced the overall and average accuracy of the model in both classification scenarios. Meanwhile, the efficiency test showed that, despite having a large number of parameters, the collaborative design of the residual structure and the dual-branch SENet significantly reduced the computational cost. The training time is only 45 ms·epoch^-1, meeting the efficiency requirements of practical combat.更多还原