In the process of image processing in frequency-domain optical coherence tomography, recursive neural networks are mainly used to achieve feature fusion. During the operation process, there is a gradient vanishing situation, which leads to low MAP (average accuracy) of multi-scale feature fusion results. Therefore, a multi-scale feature fusion method for frequency domain optical coherence tomography images based on CNN (Convolutional Neural Network) is proposed. Establish a network architecture for denoising scanned images based on generative adversarial networks, and generate high-quality scanned images without noise information through the original image domain. Using the principle of discrete wavelet transform, the denoised image is decomposed into multiple sub images. By constructing a grayscale gradient co-occurrence matrix, multi-scale image feature vectors are extracted. Starting from the local and global contrast of the image, calculate the image adaptive adjustment coefficient to achieve image detail feature enhancement processing. Finally, a feature fusion model is constructed based on convolutional neural networks, and multi-scale feature fusion results are obtained through matching analysis and concatenation processing of enhanced features. The experimental results show that after the application of the new research method, the MAP value of the multi-scale feature fusion results of frequency domain optical coherence tomography images is higher than 0.8, proving that it can effectively fuse features of different scales.