A dark and bright channel prior guided deep network for retinal image quality assessment

Retinal image quality assessment is an essential task for the diagnosis of retinal diseases. Recently, there are emerging deep models to grade images. However, current either directly transfer classification networks originally designed natural images or introduce extra priors via multiple CNN branches independent CNNs. The purpose this work address by a simple model. We propose dark and bright channel prior guided network named GuidedNet. It introduces into without parameters increasing allows training end-to-end. In detail, embedded start layer improve discriminate ability features. Moreover, we re-annotate new dataset called RIQA-RFMiD further validation. proposed method evaluated on public Eye-Quality our re-annotated RIQA-RFMiD. obtain average F-score 88.03% 66.13% RIQA-RFMiD, respectively. investigate utility assessment. And GuidedNet embedding CNNs much model burden. valid GuidedNet, re-create With achieves state-of-the-art performances