ABSTRACT Radar signals classification plays an increasingly significant role in electronic reconnaissance systems. The convolutional neural networks (CNN) with its rapid development and advancement has achieved unparalleled performance in this area. However, the recognition accuracy of CNN models is greatly suppressed due to the destruction of noise to signal features, as well as the sensitivity of these models to noise disturbances. To address this dilemma, after exploration and extensive comparative experiments, the authors found that the influence of noise ultimately maps to the pixel values of the final output feature maps, which in turn affects the final classification accuracy. To mitigate this interference, the authors propose the concept of compressing the output feature maps, and correspondingly provide a simple and efficient compression approach, which has achieved significant accuracy improvement in multiple typical CNN models (ResNet18, ResNet50, GoogLeNet, and MobileNetV2). The improved models not only avoid additional parameters or complex structures but also provide a design guideline, significantly reducing trial‐and‐error efforts. The main objective of this paper is to convey a design concept which can optimise any previously unexplored models in this topic, rather than provide a state‐of‐the‐art model.
Guo et al. (2026) studied this question.