Self-attention and Online Hard Example Mining Based Network for Marine Microalgae Detection

Abstract

With the utilization and exploitation of marine resources, the consciousness of protecting the environment is rising, and the classification and localization of marine microalgae is a good solution. In this regard, we propose self-attention and online hard example mining based network for marine microalgae detection, which is based on Cascade-RCNN network. First, the Mixup method is introduced to enhance and augment data. In the backbone network, Transformer self-attention and feature pyramid network (FPN) are introduced to make the model getting stronger feature extraction ability and can adapt to objects of multi-scale. By introducing online hard example mining (OHEM) method, the training can be completed under the condition of imbalanced data distribution. We also use multi-scale training and multi-scale testing methods to improve the training performance of the model. Through experiments on the marine microalgae dataset provided by IEEE UV 2022 "Vision Meets Algae" Object Detection Challenge, compared with the baseline network, our proposed method improves by 3.97%.