Neural Network-Based Algorithm for Detecting Key Parts of Vehicles in SAR Images

In recent years, the application of neural networks in Synthetic Aperture Radar (SAR) image processing has made important progress. A large amount of work has proven that object recognition neural networks designed for optical images also can be used for SAR images. However, there is relatively little discussion on the detection of key parts of vehicle targets based on neural networks. This article further studies the detection algorithm for key parts on the basis of existing object recognition methods for SAR images. A large number of vehicle SAR images are simulated and annotated as a dataset for network training and testing. This article uses Feature Pyramid Network (FPN) as the backbone network, which is the foundation of multi-scale object recognition. Two task specialized subnetworks attached to the backbone network respectively achieve classification and locating functions. The Vehicle Classification Neural Network (VCNN) and the Vehicles' Key Parts Detection Neural Network (VKDNN) are trained separately. Every image is first processed by the VCNN, and then the decision is made on whether to input it into the VKDNN and obtain the results. In this dataset, each image takes 0.129 seconds to be detected on average, and the key parts recognition accuracy is 91.12% in the entire dataset. Good results have been obtained in terms of detection efficiency and accuracy.