Speckle noise exists in Optical Coherence Tomography (OCT) and affects the quality of OCT images. In the diagnosis of various common eye diseases by OCT equipment, high quality OCT images are extremely important. Deep neural network is used to reduce the noise of OCT images, so that the images can show more information on the basis of retaining the details of spatial structure. A novel OCT image denoising network, CMCNN, based on residual learning network, is proposed. It has the characteristics of multi-scale, multi-weight and multi-level feature fusion, and reduces image noise while preserving the spatial structure of image details. Then the proposed model is compared with traditional denoising algorithm and deep learning denoising model. Experimental results show that the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of CMCNN are improved by about 2.5% compared with other deep learning methods. It is verified that the proposed method can effectively retain the details of OCT images, suppress the noise and improve the image quality.