Intelligent Annotation Algorithm Based on Deep-Sea Macrobenthic Images

Abstract

In the field of image processing, due to the need of expertise and skills in deep-sea biology and the disadvantages of high labor cost and long time consuming, it has always been a difficult task to mark the images of deep-sea benthic organisms. To solve this problem, this paper proposes a new image intelligent labeling algorithm LACP AL (Localization-Aware-Choice and Pseudo Label Active Learning) which is based on Localization-Aware Active Learning. LACP AL is an active learning framework based on Faster R-CNN, it finds the "valuable" samples from unlabeled samples by clustering algorithm for every training; it selects hard-to-identify samples for manual annotation and further optimizes the model; and it proposes an improved pseudo-labeling mechanism to expand the training set and improve the model accuracy. According to the publicly available dataset provided by 2020 China Underwater Robot Professional Contest, a series of experiments has been done to verify that our algorithm can achieve higher recognition accuracy with fewer training samples compared with the existing algorithms for Marine benthic image recognition.