SimpleCopy: A Strong Data Augmentation for Microalgae Detection

Abstract

Marine microalgae detection is of great importance to the environment and ecosystem. In this paper, we consider microalgae detection as a computer vision task and use a two-stage object detection network, Cascade R-CNN, to build our detector and deal with the dataset which contains a variety of small targets. Firstly, We proposed a novel data augmentation strategy called SimpleCopy for microscopic images, which typically have more small targets and sparse target distributions. Secondly, we leverage the strengths of different backbone and employ model ensemble techniques to enhance the performance of our detector. Finally, with carefully designed post-processing methods, the recall and precision of our detector can be further improved. Extensive experiments conducted on the marine dataset show the superiority of our model. We verified the effectiveness of our method by achieving 58.18 mAP and ranked 3/347 on the official leadboard.