结合卷积注意力机制改进YOLOv5s的垃圾检测

Improved YOLOv5s for Garbage Detection Combined with Convolutional Block Attention

针对传统生活垃圾检测模型检测精度低,假阳性和假阴性高的问题,提出一种结合卷积注意力机制改进YOLOv5s的垃圾检测算法。首先,利用改进的双阈值非极大抑制算法(NMS)查找原始YOLOv5s的锚框中置信度较高的检测框;然后,利用卷积注意力机制(CBAM)改进CSPDarknet53特征提取网络,强化映射到深度特征空间的特征的表达能力;最后,在自制垃圾检测数据集上对改进的网络进行训练,实现垃圾的快速定位与识别。通过在真实场景中进行测试,结果表明所提模型可以实现不同形态的多种垃圾定位与识别,平均识别精度达到95.61%,召回率达到94.85%,F1值可以达到95.70%,同时可以实现单幅图像6.01 ms的检测时间开销,满足实际应用需求,有助于促进垃圾智能化检测的效率。

Targeting at the problem of the low detection precision and high false positive and false negative of traditional garbage detection model,an improved YOLOv5s for garbage detection method combined with CBAM is proposed.Firstly,the improved dual-threshold non-maximum suppression(NMS)algorithm is used to find the detection frame with high confidence in the anchor frame of original YOLOv5s.Then,the convolutional block attention module(CBAM)is used to improve the feature extraction network of CSPDarknet53,and enhance the expression ability of features mapped into the deep feature space.Finally,the improved network is trained on the self-built garbage detection dataset to achieve the fast location and identification of the garbage.The results of testing in real scenes show that the proposed model can achieve the location and identification for various garbage of different forms,with the average precision reaching 95.61%,recall reaching 94.85%,and F1 value reaching 95.70%.Meanwhile,it can achieve the detection time overhead of 6.01 ms for a single image,and the requirements of practical application is fufilled.It is helpful to promote the efficiency of intelligent garbage classification.