In the forensic investigation, recognizing and locating the multiple diatom objects in an image is a challenging issue due to the interferences of the highly complex backgrounds. To address this issue, a coarse to fine diatom recognition and a localization framework based on the deep learning network is proposed in this paper. Firstly, the diatom images are obtained by performing the anatomic study on the cadavers. Next, a high definition electron microscope is scanned. Then, a coarse to fine deep learning framework is constructed to locate and recognize the diatom objects. Unlike the existing diatom classification and recognition methods, which used light microscopy with low resolution and completed under a simple backgrounds, our framework utilizes the high definition electron scanning microscopy with much higher resolution and suffers from the complex backgrounds interferences. To demonstrate the effectiveness of the proposed framework, 4 diatom image datasets with different background interference degrees are constructed. Also, 3 computer numerical simulation analysis are performed. They are (1) the limitations of the traditional methods in the diatom recognition, (2) the optimized composition of the training strategies and the network models, and (3) the performance of the proposed framework. The computer numerical simulation results show that the proposed framework achieves a recognition accuracy of 0.852. This is greater than 0.758 achieved by the AlexNet. Moreover, it can overcome the problem of the highly complex backgrounds interferences in the forensic investigation. Furthermore, it can locate and recognize the multiple objects in various diatom images simultaneously.

在法医调查中，由于高度复杂背景的干扰，识别和定位图像中的多个硅藻对象是一个具有挑战性的问题。为了解决这个问题，本文提出了一种基于深度学习网络的粗到细硅藻识别和定位框架。首先，通过对尸体进行解剖研究获得硅藻图像。接下来，扫描高清晰度电子显微镜。然后，构建一个由粗到细的深度学习框架来定位和识别硅藻对象。不同于现有的硅藻分类识别方法，采用光学显微镜，分辨率低，在简单的背景下完成，我们的框架利用高分辨率电子扫描显微镜，并受到复杂背景干扰的影响。为了证明所提出框架的有效性，构建了4个具有不同背景干扰程度的硅藻图像数据集。此外，还进行了3次计算机数值模拟分析。它们是（1）传统方法在硅藻识别中的局限性，（2）训练策略和网络模型的优化组合，以及（3）所提出框架的性能。计算机数值模拟结果表明，所提出的框架实现了0.852的识别精度。这大于 AlexNet 实现的 0.758。而且，它可以克服法医调查中高度复杂的背景干扰问题。此外，它可以同时定位和识别各种硅藻图像中的多个物体。