Small Target Classification of Maritime Radars Using Standardized Time-Frequency Distribution Images and Residual Networks With Improved Focal Loss

Small target classification is an important but difficult task in maritime surveillance radar measurements. Due to the small sizes of small targets, no spatial scattering structural information is available for classification. The radar cross section (RCS) fluctuation and radial velocity variation of small targets form the main information of their classification. In this article, a small target classification method based on standardized time-radial velocity images and residual network with improved focal loss (ResNet-IFL) is proposed to classify four types of small targets. The proposed preprocessing including the coordinate, size, and intensity standardization transfers a time-frequency distribution (TFD) of target returns to a standardized time-radial velocity image conveying target RCS fluctuation and movement patterns. The generalizability of the classification method includes that data from different radar measurements can be used to train classifiers and the trained classifiers can also be applied to other maritime radar measurements. Using the standardized images as inputs, classifiers are built by ResNet-IFL. Different from the focal loss that puts more attention on difficult samples, the improved focal loss uses a sigmoid-shaped weight function and puts more attention on middle samples. The proposed classifiers are used to classify the four classes of small targets, anchored buoys, floating boats, moving small targets, and low-altitude autonomous aerial vehicles (AAVs) from the IPIX, CSIR databases, and two datasets using AAV as test targets in oceanic background. Experimental results show that the classifiers based on the residual network (ResNet) and improved focal loss attain better and more robust performance than the classifiers based on other neural networks.