一种面向嵌入式平台的轻量化垃圾检测算法

A Lightweight Garbage Detection Algorithm for Embedded Platforms

摘要:

生活垃圾及其危害已引起人们的关注，而机器人与目标检测技术的发展为生活垃圾的自动化处理带来了可能性；针对目前生活垃圾检测算法在背景复杂、目标尺寸多样的情况下检测精度低，模型参数量大，深度学习检测算法综合性能不平衡以及在嵌入式设备难以部署等问题，提出了一种改进YOLOv5的轻量化垃圾检测算法；在YOLOv5模型中用GSConv模块代替传统卷积降低计算复杂度，引入了CBAM注意力机制，以提取和融合空间和通道信息，增强了网络对目标的表达能力，通过权重量化将模型进行压缩以减少模型大小加快推理速度；实验结果表明，相比于原始的YOLOv5算法，改进算法在模型的准确率和平均精确度分别提高了3%和2.3%，文件大小减小了26.6%，综合性能超越了传统的深度学习目标检测算法，对嵌入式平台更加友好。

Abstract:

The issue of domestic waste and its associated hazards has captured people's attention. The advancements in robotics and object detection technologies have opened up prospects for the automated processing of domestic waste. In light of the problems such as the low detection accuracy of current domestic waste detection algorithms in complex backgrounds and with diverse target sizes, the large number of model parameters, the imbalance in the comprehensive performance of deep learning detection algorithms, and the challenges in deployment on embedded devices, an lightweight garbage detection algorithm based on improved YOLOv5 has been proposed. In the YOLOv5 model, the GSConv module is employed to substitute the traditional convolution to lower the computational complexity. The CBAM attention mechanism is introduced to extract and fuse spatial and channel information, thereby strengthening the network's expressive capacity for the target. The model is compressed via weight quantization to reduce the model size and accelerate the inference speed. Experimental outcomes indicate that in comparison with the original YOLOv5 algorithm, the improved algorithm has raised the accuracy and average precision of the model by 3% and 2.3% respectively, reduced the file size by 26.6%, and its comprehensive performance exceeds that of traditional deep learning object detection algorithms, presenting greater friendliness to the embedded platform.