A Lightweight Underwater Object Detection Model: FL-YOLOV3-TINY

Abstract

Due to the variety of underwater object species and small object, the traditional object detection model is difficult to adapt to underwater object detection in accuracy and real-time. In this paper, a lightweight detection model FL-YOLOV3-TINY is proposed, which improves the detection accuracy and real-time performance while shrinking the model size. In FL-YOLOV3-TINY, first, the model reduces the number of parameters by introducing deep separable convolutional module to replace traditional convolutional feature extraction module. Secondly, in order to improve the detection ability of small objects and obtain more delicate image features, FL-YOLOV3-TINY adds the feature size to the three-scale to improve the detection performance. Finally, the CIoU loss regression function is introduced to make the prediction box closer to the actual box. Experiments show that compared with other lightweight models YOLOV3-MobilenetV1 and YOLOV3-Tiny, FL-YOLOV3-TINY has better mAP performance (13.7% and 10.9% increase, respectively) and better real-time perfurmance(6% and 29% increase in FPS, respectively). Meanwhile, the model size is reduced by 43% compared to YOLOV3-Tiny.