Optimizing Real-Time Object Detection Neural Networks for Dead Fish Detection in Aquaculture Systems

Detection of dead fish in the ponds is beneficial from the standpoint of interrupting the disease transmission and improving the water quality in the culture units. Nonetheless, manual and conventional methods of image processing acquisition are usually slow, less effective, and are greatly constrained by environmental factors such as water turbidity and lighting. To resolve these issues, an optimized real-time object detection neural network that follows the You Only Look Once version 4 (YOLOv4) design is therefore presented for detecting dead fish. This solution uses Mosaic augmentation for increasing dataset variability, and DropBlock regularization for decreasing over fitting and increasing model’s performance in different conditions. According to the proposed model, the global identification accuracy stands at 92 percent or 12 percent better than traditional models. In addition, these enhancements result to reduction of annotation requirements by 35% and over fitting by 20%. Compared to other related works published in recent years for object detection in aquaculture systems, the present approach shows improved results in terms of both efficiency and capacity, which confirms that this is a highly effective tool for real-time dead fish detection.