The electromagnetic environment of the working frequency band of High-Frequency Surface Wave Radar(HFSWR) is extremely complex. Noise, including radio frequency interference, sea clutter, and ionospheric clutter, can severely affect the accuracy of ship target identification. To address this issue, an improved feedforward Denoising Convolutional Neural Network(DnCNN) is proposed to suppress the noise in HFSWR marine echo signals. Based on the characteristics of the noise in HFSWR marine echo signals, the original DnCNN is modified in terms of patch size, convolutional kernel size, and network depth to make it suitable for the HFSWR denoising task. A dataset containing 10 000 pairs of Range-Doppler(RD) spectra is generated based on HFSWR sea trial data and is evenly divided into training and testing sets. Analysis of the denoising results of three groups of RD spectra in the testing set(with sea clutter, radio frequency interference, and ionospheric clutter as the main noise sources, respectively) shows that the improved DnCNN model significantly outperforms the original DnCNN in both noise suppression and maintaining the amplitude of ship target signals. Moreover,statistical results of the entire testing set indicate that the Peak Signal-to-Noise Ratio(PSNR) of the improved DnCNN denoising metric is 44.13 dB on average, which is significantly higher than the 35.58 dB of the original DnCNN. In summary, the improved DnCNN effectively suppresses the noise in HFSWR marine echoes while well preserving the amplitude of ship target signals.