A Deep Learning Method With Wavelet Packet Transform for Infrared Target Recognition
Bingzhi Yu, Huanzhang Lu, Huamin Tao · 2021
Space object recognition is an important process in infrared (IR) monitoring system. The gray value time series of the IR detector is the only data available for recognition due to long observation distance. Besides, the non-uniformity of the detector pixels and imaging noise make it difficult to extract features for recognition. In this paper, a wavelet packet transform convolution neural network (WPCN) is proposed for feature learning and classification. Different from other deep learning and handcrafted feature based methods, our method use the wavelet packet transform (WPT) to extract time-frequency domain information as the preprocessing of signal. Then the convolution neural network (CNN) extracts high-level features of decomposition data to predict label. WPCN introduces residual learning block to build the network making the learning performance improved. Training data are generated from IR gray value simulation model combined with micro-motion dynamics, geometry characteristics of space targets and imaging effect of IR detector. We evaluate WPCN on three datasets obtained by 20, 40 and 60 Hz sampling frequency. The results indicate that our method promotes the performance. The classification accuracy can reach 95% at 60Hz.