Generalizable Underwater Image Quality Assessment With Curriculum Learning-Inspired Domain Adaption

The complex distortions suffered by real-world underwater images pose urgent demands on accurate underwater image quality assessment (UIQA) approaches that can predict underwater image quality consistently with human perception. Deep learning techniques have achieved great success in many applications, yet usually requiring a substantial amount of human-labeled data, which is time-consuming and labor-intensive. Developing a deep learning-based UIQA method that does not rely on any human labeled underwater images for model training poses a great challenge. In this work, we propose a novel UIQA method based on domain adaption (DA) from a curriculum learning perspective. The proposed method is called curriculum learning-inspired DA (CLIDA), aiming to learn an robust and generalizable UIQA model by conducting DA between the labeled natural images and unlabeled underwater images progressively, i.e., from easy to hard. The key is how to select easy samples from all underwater images in the target domain so that the difficulty of DA can be well-controlled at each stage. To this end, we propose a simple yet effective easy sample selection (ESS) scheme to form an easy sample set at each stage. Then, DA is performed between the entire natural image set in the source domain (with labels) and the selected easy sample set in the target domain (with pseudo labels) at each stage. As only those reliable easy examples are involved in DA at each stage, the difficulty of DA is well-controlled and the capability of the model is expected to be progressively enhanced. We conduct extensive experiments to verify the superiority of the proposed CLIDA method and also the effectiveness of each key component involved in our CLIDA framework. The source code will be made available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/zzeu001/CLIDA</uri>.