Abstract: Radar signal modulation recognition plays a crucial role in acquiring information from non-cooperative sources and serves as a key basis for threat assessment in electronic reconnaissance systems. Traditional approaches for this task have typically faced two significant challenges： unsatisfactory recognition accuracy under low signal-to-noise ratio conditions and the prohibitively high computational complexity of conventional convolutional neural networks， which limits their practical deployment in resource-constrained environments. To address these limitations， this paper proposes a novel radar signal denoising and recognition method that integrates a hybrid attention mechanism with an improved lightweight MobileNetV4 network. The approach initially applies Choi-Williams time-frequency distribution analysis to transform 13 distinct types of radar signals into their time-frequency representations. These representations subsequently undergo grayscale conversion and normalization procedures to prepare them for network processing. Capitalizing on the inherent sparsity of radar signals in the time-frequency domain， we developed a denoising network that incorporates dedicated channel and spatial attention modules to extract critical features， combined with an enhanced U-Net architecture for effective noise suppression. The lightweight MobileNetV4 network is simultaneously optimized for Time-Frequency Representations（TFRs） characteristics through a custom-designed time-frequency aware frontend with multi-directional convolutional kernels and embedded attention mechanisms during feature extraction， significantly strengthening its perception of time-frequency structures. The denoising and recognition models were organically integrated to form an end-to-end signal processing pipeline. Simulation experiments demonstrated that the proposed denoising network substantially improved both peak signal-to-noise ratio and structural similarity index compared to original noisy TFRs. The improved MobileNetV4 recognition network achieved superior performance under low-SNR conditions， reaching 94.9% accuracy at -10 dB while maintaining only 2.57 million parameters， outperforming comparable models in both lightweight design and robustness. Experimental results confirmed that incorporating the denoising preprocessing stage effectively enhanced recognition performance in low-SNR environments. This research provides an efficient and practical solution for radar signal recognition in challenging electromagnetic environments.