There are Plenty of Fish in the Sea- Improved Optical Visibility Range for Automatic Fish Surveys and Monitoring

Optical imaging can provide indispensable information for fisheries, such as fish abundance, size, health and species classification in high spatial and temporal resolution. Nevertheless, its actual real-world use is challenged due to limited optical visibility. When aiming to train deep-learning methods for automatic fish detection, an additional challenge is that deep networks learn the appearance of the turbidity conditions of the available data, and their performance is compromised on other conditions. This is especially undesired as anyway annotating images or video is an onerous task and the available data is scarce. To tackle these challenges, we present a deep learning framework for automatic fish detection combined with an underwater visibility enhancement algorithm. First, we established a deep learning framework that is able to learn from a relatively limited set of annotated data. Second, we developed a method for increasing the visibility range in underwater images that can double the range and recover missing details in the images. We use the results of our visibility enhancement method on the training data as a data augmentation method, effectively increasing the training dataset and making it more robust to different visibility conditions. The test data is also enhanced before processing. Combined, our framework is significantly more efficient in fish detection than previous methods, and in harsher (turbid) conditions. We expect that this methodology could greatly improve fisheries science, monitoring activities and data gathering.