Design and Implementation of DRFCN deep network for military target identification

Automatic target recognition (ATR) is a process of identifying and classifying target from detected information by using automatic data processing methods. According to the levels of recognition, it can be divided into three categories: target detection, target classification and target identification. Automatic target recognition is the basis for decision making, real-time command and decisions. The combination of automatic target recognition and target state estimation is the basis for battlefield situation assessment and threat estimation. Automatic target recognition is an important component of modern weapons and equipment, and it is highly valued by academic and application departments in various countries. In recent years, with the advance of deep learning technology, the ATR technology has been extended from theoretical exploration and experimental simulation to practical application. Research on automatic target recognition technology based on artificial intelligence and deep learning has become a hot research area and frontier topic in the military field.

  Professor Liu Jun, from the information fusion team of Hangzhou Dianzi University, Information Communication and Integration National Defense Key Science Laboratory, proposed a new DRFCN deep network model for military target recognition applications. The overall structure of the model is shown in Figure 1. Different from the traditional one-way connection modeling method, the model is densely connected by convolution modules. Each convolution module is densely connected with all convolution modules of the upper layer and the characteristics of each layer in the deep network model are reused. The underlying convolution module learns all the convolution modules above. The characteristics of the feature network enhance the feature representation ability of the deep network model. After constructing the DRFCN model, the DRFCN-based scalable real-time military target recognition prototype system is designed and implemented based on the Caffe framework. As shown in Figure 2, the system consists of several functional modules such as military target recognition and military target recognition in the video. Using the developed military target recognition prototype system, the DRFCN model was verified based on the identification of battlefield military targets. The identification framework consists of nine categories: fighters, tanks, helicopters, warships, guns, missiles, cannons, submarines, soldiers. Deep network models participating in the comparative experiments include VGG, YOLO, RFCN, and ResNet. The experimental verification is compared mainly by positioning accuracy, average accuracy, false negative rate, false positive rate, network model size and processing frame rate. The experimental validation dataset includes the VOC2007&VOC2012 public dataset and a self-built military target identification dataset of the same size. As shown in Figure 3, the experimental results show that the DRFCN model can not only improve the recognition accuracy and reduce the depth network model, but also effectively solve the gradient dispersion and gradient expansion problems. In practical applications, the DRFCN model can satisfy accuracy and real-time requirements for military target identification.

Fig. 1  General structure diagram of DRFCN depth network model

Fig. 2  Demo shot of the military target recognition function based on DRFCN

Fig. 3  Partial military target recognition results

About team
Information fusion team in  Information Communication and Integration National Defense Key Science Laboratory, Hangzhou Dianzi University,  has 15 core members, all of whom have doctoral degrees. The team leader is Professor Xue Anke, and the other members include 6 professors, 6 associate professors, and 3 lecturers. The team mainly researches basic theory, technology and engineering application in the area of multi-source information fusion, focusing on three research directions, detection, tracking and recognition, image processing and image fusion and high-level information fusion. The team has published more than 500 papers and authorized more than 200 patents. During the "12th Five-Year Plan" period, it undertook 2 sub-projects of the National Basic Research Project "973", 1 national key science project and major instrument special project, 1 national defense basic scientific research project and 5 research projects from the General Armament Department. Since the "13th Five-Year Plan", the team has undertaken one sub-project of major national defense basic scientific research projects, 3 research projects of the equipment development department and 2 naval innovation projects. The team was awarded the second prize of the National Science and Technology Progress Award, the first prize of the Provincial Science and Technology Progress Award, the Provincial Science First Prize, the National Defense Science and Technology Progress Award and the Military Industry Group National Defense Science and Technology Award.

Liu Jun, Meng Weixiu, Yu Jie, et al. Design and implementation of DRFCN in-depth network for military target identification [J]. Opto-Electronic Engineering, 2019, 46(4): 180307.