中心博士生万钧的工作——“基于卷积神经网络和光子模数转换器的高精度自动目标识别”的相关成果近期被Optics Letters期刊接收，该工作得到了国家重点研发计划（2019YFB2203700）、国家自然科学基金（61822508）的部分资助。该工作研究了一维像距离分辨率与识别准确率的对应关系，得出了一维像距离分辨率的提升能提高自动目标识别准确率，由此使用具有超大带宽的光子模数转换器代替传统电子样本采集装置，并针对一维像的特点设计了优于多种深度学习算法的卷积神经网络结构用于自动目标识别。结合卷积神经网络的优势和光子模数转换的优势，提出基于卷积神经网络和光子模数转换器的自动目标识别方案，识别性能比已有多数电子方案水平高十个百分点以上。
摘要: We propose a high-accuracy automatic target recognition (ATR) scheme based on a photonic analog-to-digital converter (PADC) and a convolutional neural network (CNN). The adoption of the PADC enables wideband signal processing up to several giga-hertz and thus high-resolution range profiles (RPs) are attained. The CNN guarantees high recognition accuracy based on such RPs. With four centimeter-sized objects as targets, the performance of the proposed ATR scheme based on the PADC and CNN is experimentally tested in different range resolution cases. The recognition result reveals that high-range resolution leads to high accuracy of ATR. It is proved that when dealing with centimeter-sized targets, the ATR scheme can acquire a much better recognition accuracy than other RP ATR solutions based on electronic schemes. Analysis results also show the reason why higher recognition accuracy is attained with higher-resolution RPs.