融入注意力机制的轻量化可回收垃圾检测方法

目的 针对目前智能垃圾分类设备使用的垃圾检测方法存在检测速度慢且模型权重文件较大等问题，提出一种基于YOLOv4的轻量化方法，以实现可回收垃圾的检测。方法 采用MobileNetV2轻量级网络为YOLOv4的主干网络，用深度可分离卷积来优化颈部和头部网络，以减少参数量和计算量，提高检测速度；在颈部网络中融入CBAM注意力模块，提高模型对目标特征信息的敏感度；使用K-means算法重新聚类，得到适合自建可回收数据集中检测目标的先验框。结果 实验结果表明，改进后模型的参数量减少为原始YOLOv4模型的17.0%，检测的平均精度达到96.78%，模型权重文件的大小为46.6 MB，约为YOLOv4模型权重文件的19.1%，检测速度为20.46帧/s，提高了约25.4%，检测精度和检测速度均满足实时检测要求。结论 改进的YOLOv4模型能够在检测可回收垃圾时保证较高的检测精度，同时具有较好的实时性。

Lightweight Recyclable Garbage Detection Method Incorporating Attention Mechanism

The work aims to propose a lightweight method based on YOLOv4 to detect recyclable garbage, so as to address the problems of slow detection speed and large model weight files in the current garbage detection methods used by smart garbage sorting devices. The MobileNetV2 lightweight network was used as the backbone network of YOLOv4 and the depth-separable convolution was used to optimize the neck and head networks to reduce the parameters and computation to accelerate detection. The CBAM attention module was incorporated into the neck network to improve the sensitivity of the model to the target feature information. The K-means algorithm was used to re-cluster to get suitable self-built recyclable data with a priori frame for focused detection of targets. The experimental results showed that:the parameters were reduced to 17.0% of the original YOLOv4 model. The detected mAP reached 96.78%. The model weight file size was 46.6 MB, which was about 19.1% of the YOLOv4 model weight file. The detection speed was 20.46 frames/s, which was improved by 25.4%. Both the detection accuracy and the detection speed met the real-time detection requirements. The improved YOLOv4 model can guarantee high detection accuracy and good real-time performance in detection of recyclable garbage.