基于多分辨率特征融合的葡萄尺寸检测

Grape size detection based on multi-resolution feature fusion

针对葡萄特征提取不够充分且果粒排列密集相互遮挡难以准确检测的问题,以陕西省鄠邑区户太8号葡萄为研究对象,提出一种基于特征金字塔网络(FPN)特征融合的Faster R-CNN卷积神经网络模型完成复杂背景情况下葡萄果粒的检测与识别。以ResNet50为主干网络,引入金字塔结构,增强网络模型对葡萄果粒不同分辨率特征的提取能力,同时加入GA-RPN网络生成自适应锚框,引入遮挡补偿机制,以解决密集葡萄果粒存在的遮挡问题。模型验证结果表明,本研究提出的模型精度均值(AP)在候选框与原标记框的重叠率(IOU)阈值为50时可达95.9%,对葡萄果粒、果穗的检测准确率分别为95.8%、96.1%,相比于原始Faster R-CNN模型识别性能更优。利用双目视觉算法对葡萄果粒进行尺寸测量,在最佳测量距离(0.6~1.4 m)其相对误差可控制在2%以内。

In view of the insufficient feature extraction of grapes and the difficulty in accurate detection of dense occlusion of grape grains,a Faster R-CNN convolutional neural network model based on feature pyramid network(FPN)feature fusion was proposed to detect and identify grape grains under complex background in Huyi District of Shaanxi Province.The ResNet50 was used as the main network,and the pyramid structure was introduced to enhance the extraction ability of the network model for different resolution characteristics of grape fruit.At the same time,the GA-RPN network was added to generate the adaptive anchor frame,and the occlusion compensation mechanism was introduced to solve the occlusion problem of dense grape fruit.The model verification results showed that the average precision(AP)of the proposed model could reach 95.9%when the threshold of overlap rate between candidate box and original marked box(IOU)was 50,and the detection accuracy of grapes and strings was 95.8%and 96.1%,respectively,which was better than that of the original Faster R-CNN model.Using binocular vision algorithm to measure the size of grape seeds,the relative error can be controlled within 2%at the optimal measurement distance(0.6-1.4 m).