王祥,舒军,雷建军,杨莉.基于改进YOLACT++的碧根果图像实例分割模型[J].中南民族大学学报自然科学版,2022,41(5):613-622
基于改进YOLACT++的碧根果图像实例分割模型
Instance segmentation model of pecan image based on improved YOLACT++
  
DOI:10.12130/znmdzk.20220516
中文关键词: 实例分割  碧根果检测  Res2Net模块  CIoU损失函数
英文关键词: instance segmentation  pecan detection  Res2Net module  CIoU loss function
基金项目:国家自然科学基金资助项目(61601176)
作者单位
王祥 湖北工业大学 电气与电子工程学院武汉 430068 
舒军 湖北工业大学 电气与电子工程学院武汉 430068 
雷建军 湖北第二师范学院 计算机学院武汉 430205 
杨莉 湖北第二师范学院 计算机学院武汉 430205 
摘要点击次数: 83
全文下载次数: 88
中文摘要:
      针对碧根果加工生产线上存在果壳掉落、堆叠及打光阴影造成碧根果图像检测精度低的问题,提出一种改进的YOLACT++实例分割算法.在主干网络中采用引入了注意力机制的Res2Net模块用于增强主干网络的特征提取能力,抑制无效背景信息的干扰;在边界框回归损失函数中引入CIoU损失函数,更精确地评价预测框与真实框的位置关系,用于提高预测框的检测精度;将DIoU与Fast NMS结合,加强对重叠度高的候选框的筛选能力,改善预测框误检的问题.碧根果数据集上的实验表明:相比改进前,该算法的掩膜与预测框mAP分别提升5.18%、5.49%.COCO数据集上的实验结果表明:改进的算法对于不同尺寸物体分割精度优于BlendMask等先进算法.
英文摘要:
      Aiming at the problem of low detection accuracy of pecan image caused by falling shells, stacking, lighting and shadowing in the production line, an improved YOLACT++ instance segmentation algorithm is proposed. The Res2Net module with the introduction of the attention mechanism is used in the backbone network to enhance the feature extraction ability of the backbone network and suppress the interference of invalid background information. The CIoU loss function is introduced into the boundary box regression loss function to evaluate the position relationship between the prediction frame and the real frame more accurately to improve the detection accuracy of the prediction frame. DIoU is combined with Fast NMS to strengthen the screening ability of the candidate box with high overlap, and solve the problem of missed detection and false detection of the prediction frame. Experiments on the pecan fruit dataset show that the mask and prediction frame mAP of the algorithm are increased by 5.18% and 5.49% respectively, compared with data before the improvement. The experimental results on the COCO dataset show that the improved algorithm has better segmentation accuracy than advanced algorithms such as BlendMask for objects of different size.
查看全文   查看/发表评论  下载PDF阅读器
关闭