基于M-YOLOv4模型的轻量化目标检测算法

Lightweight target detection algorithm based on M-YOLOv4 model

针对当前网络模型复杂,网络运行速度慢、实时性存在偏低的问题。通过使用不同的策略,借鉴不同的网络模型,最终提出轻量化的模型M-YOLOv4。其在提升检测速度的同时针对小目标物体检测也有一定的提高,该方法首先引入深度可分离卷积,在保证模型精度的情况下,结合注意力机制,在实现模型的加速的同时,增强对多尺度目标的特征提取,利用各种轻量级网络的网络结构,对YOLOv4主干部分进行改进,使其对图像的局部特征进一步加强,实现模型的压缩和加速,最后通过比较各个模型的准确率、参数量及运算量证实对目标检测算法的轻量化。实验结果表明,方法针对包含小物体的不同的复杂环境,能够实现多尺度目标的检测,极大加快了模型的运行速度,且准确率达到80.12%,实现了准确率与实时性的平衡。

In view of the complexity of the current network model,the network is running at full speed and low real-time performance.This paper uses different strategies and learns from different network models,and finally proposes a lightweight model M-YOLOv4.While improving the detection speed,it also improves the detection of small target objects.This method first introduces deep separable convolution.Under the condition of ensuring the accuracy of the model,combined with the attention mechanism,while achieving the acceleration of the model,it also enhances the The feature extraction of multi-scale targets,the use of various lightweight network network structures to improve the backbone of YOLOv4,to further strengthen the local features of the image,achieve model compression and acceleration,and finally compare the accuracy of each model,The amount of parameters and the amount of calculations prove the lightweight of the target detection algorithm.The experimental results can show that the method in this paper can achieve multi-scale target detection for different complex environments containing small objects,greatly accelerate the running speed of the model,and the accuracy rate reaches 80.12%,achieving accuracy and real-time balance.