One-stage CNN detector-based benthonic organisms detection with limited training dataset

Abstract

In this paper, focusing on the challenges in unique shape dimension and limited training dataset of benthonic organisms, an one-stage CNN detector-based benthonic organisms detection (OSCD-BOD) scheme is proposed. Main contributions are as follows: (1) The regression loss between the predicted bounding box and ground truth box is innovatively measured by the generalized intersection over union (GIoU), such that localization accuracy of benthonic organisms is dramatically enhanced. (2) By devising K-means-based dimension clustering, multiple benthonic organisms anchor boxes (BOAB) sufficiently exploring a priori dimension information can be finely derived from limited training dataset, and thereby significantly promoting the recall ability. (3) Geometric and color transformations (GCT)-based data augmentation technique is further resorted to not only efficiently prevent over-fitting training but also to significantly enhance detection generalization in complex and changeable underwater environments. (4) The OSCD-BOD scheme is eventually established in a modular manner by integrating GIoU, BOAB and GCT functionals. Comprehensive experiments and comparisons sufficiently demonstrate that the proposed OSCD-BOD scheme outperforms typical approaches including Faster R-CNN, SSD, YOLOv2, YOLOv3 and CenterNet in terms of mean average precision by 6.88%, 10.92%, 12.44%, 3.05% and 1.09%, respectively.