Abstract
Deep learning (DL) has emerged as a powerful and versatile technology for achieving high-accuracy automatic target recognition (ATR) using inverse synthetic aperture radar (ISAR) images. However, there is a notable lack of analysis regarding the reliability of these recognition results, bringing the risk of misidentification. To mitigate this issue, our experiments analyze the impact of individual pixels in ISAR images, collected at varying imaging bandwidths and angles, on the predictions made by a convolutional neural network (CNN), utilizing layer-wise relevance propagation (LRP) for deeper insights. Additionally, at an imaging bandwidth of 8 GHz, we experimentally assess the imaging angles of ISAR images to establish the relationship between these angles and the accuracy and reliability of recognition outcomes. Our findings indicate that, at this bandwidth and under favorable angles, the CNN accurately identifies UAV ISAR images based on scattering points in the wings and/or the center of the UAV model, confirming the reliability of the identification results. Therefore, under large imaging bandwidths like 8 GHz, selectively accepting or rejecting CNN recognition results for ISAR images at marked favorable or inappropriate angles according to network predictions and LRP outcomes can significantly reduce the risk of misidentification.
Funding Information
  • National Natural Science Foundation of China (T2225023)
  • National Natural Science Foundation of China (62205203)

This publication has 14 references indexed in Scilit: