# Real-time object detection method based on improved YOLOv4-tiny

**Authors:** Zicong Jiang, Liquan Zhao, Shuaiyang Li, Yanfei Jia  
**arXiv:** [2011.04244](https://arxiv.org/abs/2011.04244)  
**Date:** 2020-11-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2011.04244v1/

## Key points

- YOLOv4-tiny is a simplified version of YOLOv4 designed for mobile and embedded devices
- A fast object detection method based on YOLOv4-tiny has been proposed to improve real-time performance
- The method replaces CSPBlock modules with ResBlock-D modules to reduce computation complexity
- An auxiliary residual network block is introduced to extract more feature information from objects and reduce errors
- The improved YOLOv4-tiny merges the auxiliary network with the backbone network for overall structure
- Simulation results show faster object detection compared to YOLOv4-tiny and YOLOv3-tiny models, maintaining similar precision
- This enhanced version is suitable for real-time object detection tasks, providing faster and more accurate detections.

## Summary

The "You only look once v4" (YOLOv4) is a popular object detection method in deep learning. YOLOv4-tiny is a simplified version of YOLOv4 that is specifically designed for mobile and embedded devices by reducing the network structure and parameters. To further improve real-time object detection performance, a fast object detection method based on YOLOv4-tiny has been proposed. This new method introduces several key enhancements. Firstly, it replaces the two CSPBlock modules in Yolov4-tiny with two ResBlock-D modules in the ResNet-D network to reduce computation complexity. Secondly, an auxiliary residual network block is designed to extract more feature information from objects and reduce detection errors. This block uses two consecutive 3x3 convolutions to obtain 5x5 receptive fields for global feature extraction and employs channel attention and spatial attention mechanisms for effective information extraction from objects. The overall network structure of the improved YOLOv4-tiny is constructed by merging the auxiliary network with the backbone network. Simulation results show that this proposed method achieves faster object detection compared to both YOLOv4-tiny and YOLOv3-tiny models while maintaining almost the same mean value of average precision as YOLOv4-tiny. Therefore, this approach is highly suitable for real-time object detection tasks. In conclusion, this study presents an enhanced version of YOLOv4-tiny that significantly improves real-time object detection performance through modifications made to its network structure and feature extraction process resulting in faster and more accurate detections compared to previous versions.