YOLOv5-lite: a novel compression method based on channel pruning

Abstract Underwater visual perception algorithms play a critical role in high-performance applications of underwater object detection. Although deep network-based object detection models offer superior detection rates, they typically involve abundant parameters, requiring substantial computation that makes deployment on edge-end devices challenging. Therefore, it is crucial to implement lightweight optimizations of these models to ensure the success of current deep network model object-identification techniques. In this research, we apply a convolutional kernel pruning method based on the YOLOv5 object detection model. We first use the multi-objective optimization approach to determine the sparse batch normalization layer scaling factor and bound. The obtained results serve as evaluation criteria for pruning and fine-tuning the convolution kernel parameters. Finally, we use the underwater biological dataset for training and testing. The experiment demonstrated that by pruning nearly 70% of the model parameters, we can increase the model efficiency and achieve comparable performance, with only a 2% drop in accuracy.