Underwater image super-resolution and enhancement via progressive frequency-interleaved network

Underwater images usually contain severely blurred details, color distortion, and low contrast, warranting efficient methods to obtain clean images. However, most convolutional neural network-based approaches involve high computational cost, numerous model parameters, and even poor performance. Besides, the mapping from input to output is learned using a single path, ignoring the frequency domain information. To solve these challenges, we propose a novel progressive frequency-interleaved network (PFIN) for underwater imagery super-resolution and enhancement. Specifically, progressive frequency-domain module (PFDM) and convolution-guided module (CGM) constitute PFIN for effective color deviation correction and detail enhancement. PFDM that possesses global spatial attention, multi-scale residual, and frequency information modulation blocks gradually learn frequency features and explicitly compensate for detail loss. Furthermore, CGM comprising a series of convolution blocks generates discriminative characteristics to modulate in PFDM for better accommodating degraded representations. Extensive experiments demonstrate the superiority of our PFIN regarding quantitative evaluations and visual quality.