基于改进YOLACT的堆垛图像快速分割方法研究

Research on Fast Segmentation Method of Stacking Image Based on Improved YOLACT

摘要:

针对堆叠密集的堆垛货箱出现的漏检情况以及难以分割出每个货箱的精确边缘而造成的难以准确抓取的问题,对深度学习实例分割算法YOLACT进行了相应的改进。首先使用工业相机采集货箱的堆垛图像,然后利用Labelme标注图像制作数据集,并且通过数据增强方法扩充数据集。接着为了提高模型的分割准确率,分别对掩码真值和YOLACT中的原型掩码输出分支(Protonet)的预测掩码使用Canny边缘检测算子,并取二者的二值交叉熵损失作为损失函数加入到原网络中训练。最后再使用训练好的最优模型对测试集图像数据进行试验,结果表明,改进后的模型预测掩码mAP_0.5:0.95可以达到0.543,比原模型提高2.2%,同时货箱边缘的分割精度也得到了一定的提升,模型推理速度可达10.2帧/秒,可以满足精度要求和生产节拍要求。

Abstract:

The deep learning instance segmentation algorithm YOLACT was improved to solve the problem of missing detection of densely stacked packing boxes and the problem of difficult to capture accurately due to segmenting the inexact edges of each packing box. Firstly, the industrial camera was used to collect the stacking image of the packing box, and then the Labelme was used to annotate image to create the dataset, and the dataset was expanded through the data enhancement method. Then, in order to improve the segmentation accuracy of the model, the Canny edge detection operator was used for the groundtruth value and predicted mask of the prototype mask output branch (Protonet) in YOLACT, and the binary cross-entropy loss of the them was added to the original network as a loss function. Finally, the trained optimal model was used to test the image data of the test set. The results show that the improved model predicted mask mAP_0.5:0.95 can reach 0.543, which is 2.2% higher than the original model. At the same time, the segmentation accuracy of the packing box edge has also been improved to a certain extent. The model’s inference speed can reach 10.2 frames/second, which can meet the accuracy requirements and production beat requirements.