Research on Target Detection Algorithm Based on Improved YOLO

Abstract

Aiming at the problem that the target detection algorithm has poor real-time performance and high complexity for small-size object detection, this paper creates a model that can effectively improve YOLOv4-tiny deep learning. The detection idea of merging different features is adopted in the model, and the convolution module is added to the model, each parameter is adjusted to ensure the optimization of the final network structure, and then the $52 \times 52$ target prediction scale output is added, which helps to improve the recognition of small-sized targets. accuracy. Based on the original network structure, the K-means++ clustering method is used to conduct experiments on the data set of small target objects to determine a priori frame size that is more suitable for small-sized object recognition, so that the network can more quickly and accurately identify small-sized objects. The target position of the object. The test set is trained on the network before and after the improvement, and the model is deployed on the platform based on Jetson AGX Xavier for testing. The results show that the improved YOLOv4tiny model has an increase of 3.56% in the accuracy rate and 7.08% in the recall rate. The average The accuracy is improved by 2.83%.