Fast and Accurate Fish Detection Design with Improved YOLO-v3 Model and Transfer Learning
Kazim Raza, Song Hong
Abstract
Object Detection is one of the problematic Computer Vision (CV) problems with countless applications. We proposed a real-time object detection algorithm based on Improved You Only Look Once version 3 (YOLOv3) for detecting fish. The demand for monitoring the marine ecosystem is increasing day by day for a vigorous automated system, which has been beneficial for all of the researchers in order to collect information about marine life. This proposed work mainly approached the CV technique to detect and classify marine life. In this paper, we proposed improved YOLOv3 by increasing detection scale from 3 to 4, apply k-means clustering to increase the anchor boxes, novel transfer learning technique, and improvement in loss function to improve the model performance. We performed object detection on four fish species custom datasets by applying YOLOv3 architecture. We got 87.56% mean Average Precision (mAP). Moreover, comparing to the experimental analysis of the original YOLOv3 model with the improved one, we observed the mAP increased from 87.17% to 91.30. It showed that improved version outperforms than the original YOLOv3 model.