Detection of the Sinking State of Liquid Oil in Breaking Waves Based on Synthesized Data: A Behavior Process Study of Sunken and Submerged Oil

Abstract

In computer vision, pollutant detection is a highly concerning issue, and it has been widely used in the fields of pollutant identification, tracking, and precise positioning. In the ocean, oil tends to disperse into the water column as droplets under breaking waves, and it is called sunken and submerged oil. Aiming at the most difficult issue of identifying liquid submerged oil pollution, this paper proposes a method of synthesized data containing specific markers for oil detection. The Canny operator was used to remove the background of the liquid submerged oil. Then, affine transformation was applied to simulate the real situation of oil deformation. Linear mapping was presented by matrix multiplication, and translation was represented by vector addition. At last, bilinear interpolation was used to integrate the oil into the image of the laboratory pictures. In addition, this research randomly added interference information, so that the probability distribution of synthesized data was closer to the probability distribution of the real data. Then, this paper combined various methods to improve the accuracy of liquid oil detection, such as Feature Pyramid Networks, RoIAlign, difficult sample mining. Based on the above methods, 1838 images were synthesized in this paper and combined into a training set. The results show that the average accuracy of the oil detection is increased by 79.72%. The accuracy of the synthesized data method for labeled oil detection was 18.56% higher than that of oil detection without labeling. This research solves the difficulty of obtaining sunken and submerged oil images and the high cost of image annotation.