Multimedia Tools and Applications Software
SPSO-Pruner: a network pruning method on YOLOv5 for fewer categories scenarios
Abstract
In recent years, object detection methods based on deep learning have developed rapidly, but we have found that comparing with public datasets, the categories of objects to be recognized in most practical scenarios are relatively small. In such scenarios, we proposed a network pruning method inspired by the PSO algorithm, named SPSO-Pruner to improve the detection accuracy and speed, while reducing the model parameters. To better fit the application scenarios, we used the up-to-date one stage detector - YOLOv5 to adapt to the real-time requirements. Comparing with regular network prune method such as slimming prune, our SPSO-pruner can reach better accuracy with less than 50% parameters. Besides, we proposed an optimization method for confidence Loss of YOLOv5 which can balance the Precision and Recall of our model, the F1 of our method is 4% higher than the baseline model.