基于YOLOv4的小目标检测方法研究及应用

Research and application of small target detection method based on YOLOv4

为解决复杂施工场景下的小目标检测效果不佳和漏检问题,提出一种基于YOLOv4的改进算法。在检测网络中设计多尺度CAU和SAU上下文特征融合机制,利用全新的特征融合方式增强网络多尺度空间和通道信息表征,在此基础上改善网络特征融合性能。设计CSP_F跨阶段特征融合模块代替原有普通卷积块(CBL*5),防止检测网络梯度消失和网络参数计算量过大。改进模型类别损失函数并进行实验验证,其结果表明,改进算法能满足不同场景检测要求,对小目标有较好检测效果。

An improved algorithm based on YOLOv4 was proposed to solve the problems of poor detection effect and missing detection of small objects in complex construction scenarios.A multi-scale CAU and SAU context feature fusion mechanism was designed in the detection network,and a new feature fusion method was used to enhance the representation of multi-scale spatial and channel information,and on this basis,the feature fusion performance of the network was improved.The CSP_F cross-stage feature fusion module was designed to replace the original common convolution block(CBL*5),to prevent the disappearance of detection network gradient and excessive computation of network parameters.The class loss function of the model was improved and verified by experiments.The results show that the improved algorithm can meet the detection requirements of different scenes and it has good detection effects on small objects.