Abstract: An integrated deep neural network is proposed to co-optimize signal denoising and classification in response to the performance degradation of ship classification systems caused by noise interference in the marine environment. In this network, a dual-decoder architecture is employed by the denoising module which enables parallel processing of the amplitude spectrum and phase spectrum of ship signals. Building on this, three features that integrate time-frequency information are extracted from the denoised signal and fed into a recognition module with multi-scale modeling capabilities to perform ship classification. By designing a joint denoising-classification training loss function, the proposed network enables gradient sharing and coordinated parameter updates between the denoising and recognition modules. This allows the denoising process to be directly optimized for the classification task, effectively avoiding the loss of discriminative features often observed in traditional two-stage denoising-classification approaches. Experimental results on the Shipsear dataset demonstrate that a 25.73% improvement in recognition accuracy is achieved by the proposed integrated network compared to direct classification on noisy data. Furthermore, a 12.31% improvement in recognition accuracy is obtained when compared to the traditional two-stage denoising-classification method.