An Efficient and Effective Sea Turtle Detection Using Positioning Enhancement Module

Turtle detection serves as the foundational step supporting subsequent activity analysis. This task contributes to observing the behaviors by deploying underwater robots, ensuring minimal disruption to their natural activities. Challenges such as background interference, occlusion, and small object sizes weaken the lightweight feature extraction networks, and they require an additional approach that can distinguish the characteristics of the target object. We propose two enhanced positioning modules to address these challenges: dual-spatial attention and gradual receptive enhancement. The dual-spatial attention module utilizes two representations to identify essential information at the medium-level feature layer. The gradual receptive enhancement module also enhances the filter process at high-level features by emphasizing vital elements within the convolutional group representation. As a result, the proposed detector achieves high precision with low parameters and compu-tational requirements. Furthermore, adding these modules does not compromise the inference time on an edge device, which achieves a speed of 14 FPS.