Deep Learning-Based Scallop Detection in Seabed Images Using Active Learning and Model Comparison

Abstract

This study proposes a method for detecting scallops in seabed images using deep-learning based instance segmentation and active learning techniques. This method uses a mask region-based convolutional neural network (Mask R-CNN) combined with active learning to enable efficient annotation and adaptive learning in different seabed environments. A comparison with the transformer-based deformable detection transformer (Deformable DETR) model provides a detailed evaluation of the detection performance. The proposed method proves to be effective in detecting of object features while removing unnecessary background regions in noisy seabed environments. Active learning with margin sampling enhances the annotation process and creates an effective dataset from numerous seabed images. Experiments conducted on a large dataset of over 83,000 seabed images show that Mask R-CNN outperforms Deformable DETR, achieving an F-measure of 0.89 compared to 0.85. This study contributes to the field of fishery resource investigations by providing an approach for efficient learning using new data, which is crucial for maintaining accurate scallop detection systems over time.