A Decomposition-based Network for Non-uniform Illuminated Retinal Image Enhancement

The retinal images are often used in diagnosis of eye diseases. However, due to the non-uniform illumination during imaging, a considerable proportion of retinal images are of low quality, which affect doctors’ diagnosis of eye diseases and reduce the performance of downstream diagnostic systems. To address this problem, we propose a decomposition-based network to remove the non-uniform illumination to enhance the quality of the retinal images. Our network consists of two parts: the first sub-network aims to decompose an image into illumination and reflectance maps, and the second one adjusts the estimated illumination to improve the image quality. Specifically, in the image decomposition stage, we train a deep convolutional neural network to approximate the Retinex-based decomposition algorithm to estimate the illumination and reflectance. In the image enhancement stage, an encoder-decoder network structure is adopted to adjust the estimated illumination and learn the distribution of well-lit images. Finally, the adjusted illumination and the reflectance are combined to obtain the final enhanced result. To facilitate network training, we construct a new dataset that contains paired low/high quality retinal images. Experiments show that our method is able to effectively enhance the illumination while preserving the detail and color information. Furthermore, the proposed method also has advantages in improving the accuracy of blood vessel segmentation.