Context guided transformer enhanced YOLOv8 for accurate juvenile abalone detection and counting

The accurate detection and counting of juvenile abalones are essential for estimating population biomass and culture density in aquaculture. However, due to the small size, dense distribution, and frequent occlusion among individuals during the rearing period, existing detection algorithms often demonstrate low precision in identifying abalones. In this study, we introduce the Context Guided Transformer YOLO (CGT-YOLO) model to tackle the problem, utilizing You Only Look Once version 8 (YOLOv8) as the foundational model for detecting and counting juvenile abalones. Specifically, the Context Guided (CG) module is employed to down-sample the input images, thus enlarging the receptive field and preserving more target-related information, which ultimately reduces the loss. Subsequently, the Contextual Transformer (CoT) module is incorporated within the architecture to augment the model's capacity to focus on small-sized and densely overlapped targets, thereby reducing missed and incorrect detections. In addition, by constructing a small target detection layer grounded in the lower-level, finer-resolution feature representations, the model's capacity to recognize detailed information within the image is enhanced. Finally, we employ the inner complete intersection over union (Inner-CIoU) loss to facilitate model training by optimizing bounding box adjustments through a scaling factor, which accelerates convergence and further enhances accuracy. Results obtained through experimentation on the self-built abalone dataset validate how the CGT-YOLO surpasses several existing models in detecting juvenile abalones, effectively overcoming the challenges posed by individual adhesion and overlap. This demonstrates its reliability and effectiveness in practical aquaculture applications.