Detecting salmon lice in seawater using synthetic datasets

Salmon lice are one of the most serious threats to the salmon aquaculture industry. Automatic methods for monitoring the density of salmon lice larvae in seawater are paramount for taking proper response measures. Computer vision technologies are promising but require large annotated datasets to create effective detection models. This paper investigates the potential of using methods to create datasets containing synthetic images based on the combination of salmon lice masks (collected from real images) with different environment-related backgrounds. The diversity of the synthetic dataset is ensured through a spectrum of factors, including the number of synthetic images, poses, variations in brightness, contrast, and saturation of objects and background images, ‘noise’ injection, and scale. Orthogonal experiments are conducted to assess the impact of these factors. We also evaluate the effectiveness of a state-of-the-art object detector, YOLOv8, trained on created synthetic datasets. The results show that the use of synthetic datasets leads to significant improvements, up to 75 percent, in the recall of the salmon lice detection. Rotation is the transformation that had the most significant impact on the synthetic dataset. Moreover, this approach is applicable to datasets of different sizes and other detectors. The use of synthetic datasets seems a valuable tool for the detection of ‘needles in a haystack’ such as incidental salmon louse larvae free swimming in large volumes of water with variable and complex backgrounds.