Research on Detection Algorithm Based on YOLO and Lightweight Network EfficientNet

Deep learning based object detection algorithms have been applied in various fields of life. As a representative of one-stage detection algorithms, YOLO series algorithms are highly favored for their fast speed.To resolve the existing issues with complex network architecture and weak real-time performance. YOLOv3 and YOLOv4 are improved by adding the lightweight network EfficientNet, using deep separable convolution rather than regular convolution, reducing a significant number of network parameters, using transfer learning to speed up the model’s training speed, and speeding up the model’s convergence. Finally, the accuracy, number of parameters, and detection rate of the network structure before as well as after improvement are compared to ensure that the model is successful. According to the experimental findings, the number of parameters for E-YOLOv3 and E-YOLOv4 are only 1/9 and 1/7 of what they were before the enhancement, and the detection rate has increased by 4 and 3 frames/second, respectively.