基于多尺度特征注意Yolact网络的堆叠工件分拣算法

Stacking workpieces sorting algorithm based on multi-scale feature attention Yolact network

摘要:

针对非结构化场景中存在的多工件堆叠遮挡等问题,提出了基于多尺度特征注意Yolact网络的堆叠工件识别定位算法。所提算法首先在Yolact网络的掩码模板生成分支中加入多尺度融合与特征注意机制,提升网络预测堆叠工件掩码的质量,并设计了基于膨胀编码的目标检测模块,增强网络对不同尺度堆叠工件的适应能力,构建了多尺度特征注意Yolact网络。其次,利用构建的多尺度特征注意Yolact网络预测堆叠工件的掩码与边界框,并对堆叠工件掩码进行最小外接矩形生成,根据掩码边界框与掩码的最小外接矩形确定目标工件的抓取点与旋转角度。最后,基于堆叠工件识别定位算法研发了视觉机器人工件分拣系统。实验结果表明,所提模型在边界框回归、掩码预测两项任务上的识别精度均有提升,机器人工件分拣系统进行堆叠工件分拣作业的成功率达到97.5%。

Abstract:

Aiming at the problems of multi workpieces stacking occlusion in unstructured scenes, a stacked workpiece recognition and location algorithm based on multi-scale feature attention Yolact network is proposed. Firstly, the proposed algorithm adds multi-scale fusion and feature attention mechanism to the mask template generation branch of Yolact network to improve the quality of network prediction stacking workpieces mask, designs a target detection module based on expansion coding to enhance the adaptability of the network to stacking workpieces with different scales, and constructs a multi-scale feature attention Yolact network. Secondly, using the constructed multi-scale feature attention Yolact network to predict the mask and bounding box of the stacked workpieces, generate the minimum circumscribed rectangle of the stacked workpieces mask, and determine the grab point and rotation angle of the target workpieces according to the mask bounding box and the minimum circumscribed rectangle of the mask. Finally, a visual robot workpieces sorting system is developed based on the stacking workpieces recognition and positioning algorithm. The experimental results show that the recognition accuracy of the proposed model in the two tasks of bounding box regression and mask prediction is improved, and the success rate of stacking workpieces sorting by the robot workpieces sorting system is 97.5%.