Bio-inspired optimization algorithms for real underwater image restoration

Abstract

Underwater image restoration algorithms usually take into account the physical model of the acquisition medium to obtain a restored image with good quality. This way, they compensate for the specific degradations introduced by the physical medium. These models often contain several parameters, which represent the type and strength of the medium degradations (e.g., absorption, scattering, and many others). Naturally, the quality of the restored image will depend on the correct estimation of these model parameters. In this work, we propose a restoration algorithm that estimates the model parameters using bio-inspired optimization metaheuristics, whose cost function (objective function) is a No-Reference Image Quality metric (NR-IQA). In this case, Opposition-based Particle Swarm Optimization (OPSO), Repulsive–Attractive Particle Swarm Optimization (RAPSO), Artificial Bee-Colony Algorithm (ABC), Opposition-based Artificial Bee Colony (OABC), and Differential Evolution (DE) metaheuristics have been tested. Since most quality metrics are not designed for underwater scenarios, a study was carried out to choose the best metric for this type of scenario, which is used as the cost function during the optimization process. To do that, an underwater image dataset was built, containing a set of images with underwater degradations and their corresponding subjective quality scores. The subjective quality scores were obtained by performing a subjective quality assessment experiment with voluntary participants that rated the quality of the test images. The proposed study has found out that NIQE metric presents the highest SRCC value, being therefore chosen as the cost function for the optimization algorithms that estimate the physical parameters of Barros’s inverse model. The performed experiments have demonstrated the suitability of our approach for underwater image restoring, showing that among the tested metaheuristics methods the PSO and ABC are the better, which provided restored images with a good visual quality.