I. Introduction
With the development and application of millimeter wave passive imaging technology, it is of great value to study how to detect and identify dangerous targets quickly and accurately from millimeter wave images. Many methods of infrared target detection and recognition can be used for target detection and recognition in PMMW images, due to the similarity between PMMW images and infrared images in terms of low-resolution and noise interference. The common detection and recognition methods for infrared image are mainly based on traditional segmentation, feature extraction and feature matching methods. However, the resolution of PMMW images is lower than that of infrared images, which means that the detection method of the infrared image cannot be completely applied to the PMMW images and also shows the unique challenge of concealed target detection in PMMW images. Seokwon Yeom, et al, Daegu University of Korea, has proposed a multi threshold segmentation algorithm based on K-means for detecting hidden targets in millimeter wave images[1]. The method of multilevel segmentation is used in [2] to detect the dangerous target in PMMW image. In [3], the classical LBG-VQ method is used to segment the millimeter wave image and the divided target images are obtained. The targets were only detected in above methods, but not classified. The classical classification methods include Support Vector Machine (SVM) [4], k-NearestNeighbor (KNN) [5], back propagation (BP) neural network[6] and so on. In [7],To recognize millimeter wave images using geometric descriptors, it needs to extract feature vectors manually in advance, which is cumbersome. In [8], Six classifiers of Logistic Regression (LR), RL quadratic (QLR), SVM, Random Forest (RF), Extreme Random Trees (ERT), Adaboost (ADA) and two features of Haar, Local Binary Patterns (LBP) are used by authors to classify the raw images and preprocessed images, respectively. Although good performance has been achieved, the accuracy of classification still depends the selection of exact and effective features. [9] based on the contour features of the human body, the DTW algorithm is utilized to identify the human body, and the edge features need to be extracted manually. The accuracy of the recognition is easily affected by the incomplete extraction of the contour.