Content-adaptive Efficient Transformer for No-Reference Underwater Image Quality Assessment

No-reference underwater image quality assessment (NR-UIQA) is a fundamental yet challenging task in ocean engineering field. Current methodologies for NR-UIQA, particularly those employing convolutional neural networks (CNNs), commonly leverage deeply-stacked convolutional layers to capture local features associated with image quality. However, these CNNs-methods often neglect the significance of non-local information. To overcome this limitation, we propose an innovative solution: an end-to-end content-adaptive efficient transformer (CET) designed specifically for NR-UIQA. The CET comprises a multi-scale feature extraction (MFE) backbone module and an adaptive quality regression (AQR) module. This architecture allows for the adaptive evaluation of image quality by dynamically adjusting the weights and biases of fully-connected layers within the AQR module, enhancing generalizability across diverse underwater environments. Experimental findings substantiate the superiority of CET over existing methods, showcasing its state-of-the-art accuracy and efficiency on publicly available datasets.