基于多尺度特征提取与融合的垃圾分类模型研究

Research on garbage classification model based on multi-scale feature extraction and fusion

以YOLOv4算法为基础,了解YOLOv4算法的基本结构及其在垃圾分类实际应用场景中存在的问题,提出了YOLOv4算法改进策略,通过在CBM模块中设定两个卷积核来实现不同尺寸垃圾的特征提取。对改进的多尺度特征提取与融合的垃圾分类模型进行训练与验证,发现改进的YOLOv4算法可以很好地满足多尺度特征提取与融合的基本需求,对小尺寸垃圾目标的分辨效率可达85%以上,处理时间也能很好地满足实际应用场景的基本需求,并且卷积核值设定越小,垃圾目标识别精确度越高,处理时间也会相应增加,可以很好地促进垃圾分类产业向智能化转变。

Taking the YOLOv4 algorithm as the foundation,the paper proposes improvement strategies for the YOLOv4 algorithm by understanding its basic structure and addressing the issues encountered in practical applications of garbage classification.The improvement is achieved by setting two convolutional kernels in the CBM module to facilitate the feature extraction process for garbage of different sizes.Through training and validation of the improved garbage classification model with multi-scale feature extraction and fusion,it is found that the improved YOLOv4 algorithm can effectively meet the basic requirements of multi-scale feature extraction and fusion.The recognition efficiency for small-sized garbage targets can reach 85%over,and the processing time can meet the basic requirements of practical applications.Furthermore,it is observed that as the convolutional kernel size decreases,the accuracy of garbage target identification increases while the processing time also increases accordingly.This research contributes to the advancement of the garbage classification industry towards intelligence.