基于Mask R⁃CNN算法的塑料瓶实例分割方法

Plastic bottle instance segmentation method based on Mask R⁃CNN algorithm

塑料瓶是对海洋和陆地环境构成严重威胁的主要污染物之一。在复杂背景下准确、及时检测塑料瓶是塑料废弃物回收工作面临的巨大挑战。因此提出一种基于掩膜区域的卷积神经网络(Mask R⁃CNN)和迁移学习对复杂背景下塑料瓶进行检测的方法。首先对图像数据进行采集和标注,建立塑料瓶图像数据集;随后运用迁移学习将COCO数据集上的预训练网络参数迁移到塑料瓶实例分割模型中作为初始化参数,并对模型进行训练直至收敛。实验显示,未进行迁移学习的mAP为56.0%,迁移学习组的mAP为59.4%,并且在mAP_va[l 0.5:0.05:0.95]上取得了较0.9(90%)检测最小置信度提高0.45和0.17的边际改进。实验结果表明,Mask R⁃CNN模型可以较好地应用于塑料瓶图像的分割,且迁移学习对模型训练效果的提升较大。

Plastic bottle is one of the major pollutants that pose a serious threat to marine and land environment.Accurate and timely detection of plastic bottles under complex background is a great challenge for plastic waste recycling.Therefore,proposing a method based on Mask Regions with Convolutional Neural Network Features(Mask R‑CNN)and transfer learning to detect plastic bottles in complex backgrounds.Firstly,the image data was collected and labeled,and the plastic bottle image data set was established.Then transfer learning was used to transfer the pre‑training network parameters from COCO data set to the plastic bottle instance segmentation model as initial parameters,and the model was trained until convergence.The experiment shows that mAP without transfer learning is 56.0%,while mAP of transfer learning group is 59.4%,and marginal improvement of mAP_val[0.5:0.05:0.95]is 0.45and 0.17 higher than 0.9(90%)detection minimum confidence.The experimental results show that the Mask R‑CNN model can be well applied to the segmentation of plastic bottle images,and transfer learning can greatly improve the training effect of the model.