Real-Time Sea Cucumber Detection Based on YOLOv4-Tiny and Transfer Learning Using Data Augmentation

Thao NgoGia, Yinghao Li, Daopeng Jin, Jinyuan Guo, Jiang Li, Qirong Tang

Abstract: You Only Look Once version 4 (YOLOv4) model has an outstanding performance in object detection and recognition. However, the YOLOv4 is too complex, requiring high computing resources with a lot of training data, which is difficult in the underwater environment. YOLOv4-tiny is proposed based on YOLOv4 to simplify the network structure and reduce parameters, which makes it be suitable for developing on mobile and embedded devices. In this paper, in order to implement a real-time cultured sea cucumber detector to the autonomous underwater vehicle (AUV), YOLOv4-tiny and transfer learning are applied. The model has a good performance in speed but the accuracy is unsatisfactory while evaluated on the real-world underwater datasets. Therefore, a data augmentation method based on improved Mosaic data augmentation is further proposed to improve the quality of the training dataset. The proposed method is evaluated on the real-world sea cucumber underwater videos and has a good performance.