Abstract: To address the significant degradation of ship high resolution range profiles （HRRP） classification performance under low SNR caused by feature degradation， this study proposes an adaptive residual depthwise shrinkage network （ARDS-Net） that fuses handcrafted features and deep learning for ship recognition. Firstly， Total Variation （TV） denoising and Krogager polarization decomposition are performed on radar dual-polarization data to extract hybrid handc-rafted features integrating polarization statistical features and wavelet anti-noise features. Then， a residual shrinkage building unit with channel-wise thresholds （RSBU-CW） based on depthwise separable convolution is designed to suppress noise via learned dynamic soft thresholds. Finally， focal loss is used to optimize the weights of hard samples under low SNR. Experiments on a measured dataset with sea clutter and electromagnetic interference show stable anti-noise performance from -10 to 10 dB， with accuracies of 52.94% （0 dB）， 73.05% （5 dB）， and 91.11% （10 dB）. Relying on the multi-dimensional anti-noise mechanism， the proposed method effectively improves the robustness of ship radar polarization feature recognition under complex electromagnetic environments， and can provide technical support for engineering applications such as maritime security and ship monitoring.