Synthetic Data for Semantic Segmentation in Underwater Imagery

Semantic segmentation is a critical component of scene understanding in underwater environments. Deep learning models have shown the best performance in semantic segmentation tasks. Deep learning benefits from large amounts of labeled imagery for training. However, there is a significant gap in size between the above-ground and underwater datasets. This work addresses the dataset size and class imbalance problems of underwater datasets by creating synthetic images to increase effective dataset size. Using the semantic segmentation mask, we can cut and paste objects from one image into another. A synthetic dataset created this way could also address class imbalances by automatically adding more instances of underrepresented classes. Doing this can create a subpixel boundary around inserted objects that networks will learn to detect instead of the intended targets. To solve this problem, we compare two different blurring methods, Gaussian and Poisson, to reduce the impact of this effect. We evaluate different methods, such as Gaussian and Poisson blurring and the number of targets added to synthetic images, to find the best way to blend inserted objects in underwater imagery. In addition, we assess the different methods to see which one produces the most realistic training set by comparing the performance metrics of various semantic segmentation algorithms after training on synthetic datasets employing each method.