Streamlining Deep Learning Network for Real-time Sea Turtle Detection

Abstract

Monitoring turtle behavior is a conservation effort to preserve its habitat, and the detection process is a vital initial stage. On the other hand, robotics demands a deep learning network to automatically detect the presence of sea turtles that can operate in real-time. The need for increased model speed in the inference stage has led to many lightweight vision-based detectors. This work proposes a novel turtle detection to localize multiple sea turtles using a deep learning method. A lightweight primary extractor is applied to distinguish crucial features without producing a huge computational. An excited group attention is offered as an enhancement module that can capture essential turtle components in multi-level convolutional patches. A new turtle dataset is proposed that contains lighting, blur, occlusion, and complex background challenges. The evaluation results show that the proposed model performs higher accuracy than other lightweight object detection models. High-efficiency benefits models that can be implemented on low-end devices in terms of real-time data processing speed.