基于改进YOLOv4-tiny的零件目标检测

Part Target Detection Based on Improved YOLOv4-tiny

针对无序多角度零件视觉识别准确率不高,定位精度低的问题,提出一种基于YOLOv4-tiny的改进神经网络算法。改进的算法主要是将CBAM注意力机制引入到YOLOv4-tiny网络,使特征提取网络关注重要特征区域,并过滤无关信息。采用K-means算法对数据集进行聚类,重新得到anchor的对应参数。在零件数据集上进行对比实验,测试结果表明:所提算法在满足实时性的基础上,准确率相比原网络提高了3.4%,平均精确率提高了1.8%,具有较好的综合检测能力。该研究可为工业机器人的零件智能分拣提供技术参考。

Aiming at the problems of low visual recognition accuracy and low positioning accuracy of disordered multi angle parts,an improved neural network algorithm based on YOLOv4-tiny is proposed. The improved algorithm mainly introduces the CBAM attention mechanism into the YOLOv4-tiny network,which makes the feature extraction network focus on important feature regions and filter irrelevant information. The K-means algorithm is used to cluster the dataset and regain the corresponding parameters of anchor. Comparative experiments are conducted on the part dataset,and the experimental results show that the proposed algorithm has a better comprehensive detection capability by improving the accuracy rate by 3.4% and the average accuracy rate by1.8% compared with the original network while satisfying the real-time performance. This research can provide technical reference for intelligent part sorting of industrial robots.