基于改进U-Net网络的大棚草莓果实识别

Image segmentation of strawberry in greenhouse based on improved U-Net network

[目的]国内自动化果蔬采摘设备正处于起步阶段,草莓自动化采摘及品质鉴定受图像分割和识别精度直接影响。目前真实大棚种植环境中光照和浆果及叶片间遮挡对草莓的分割、识别、定位、分级带来极大的挑战。[方法]本文以真实大棚草莓为研究对象拍摄并制作数据集,为提升模型的泛化能力,采用色彩抖动(即随机进行饱和度、亮度、对比度调整)和随机旋转两种方式进行数据增强操作,共获得大棚草莓数据集14147张。以U-Net分割模型为原型进行改进设计,编码器基于VGG16模型结构进行修改,选用带same的卷积来避免Feature Map合并前的裁边操作,从而保留更多的特征信息,增加批归一化(Batch Normalization)操作来加快训练过程并解决梯度消失的问题。以此提高试验训练速度和模型训练精度,将改进后的U-Net模型与原始U-Net、PSPNet、DeeplabV3模型在同等的试验数据和参数下进行对比分析。[结果]在选取无遮挡粘连和有遮挡粘连2类测试数据集上,本文所提出的方法分割效果更优且获得了更好的鲁棒性,其识别精度为96.05%,MIOU值为89.41,F1_score值为0.9。[结论]改进后的U-Net模型在大棚草莓数据集上的分割效果最佳,其结果可为后续草莓自动化采摘和品质等级鉴定提供有力的技术支撑。

[Objective]The domestic automatic fruit and vegetable picking equipment is in its infancy,and the automatic strawberry picking and quality identification are directly affected by the image segmentation and recognition accuracy.At present,the illumination and the occlusion between the leaves and berries in the real greenhouse planting environment bring great challenges to the segmentation,identification,positioning and classification of strawberries.[Method]In this paper,the real greenhouse strawberries were used as the research object and a data set was created.In order to improve the generalization ability of the model,The methods of color dithering(random adjustment of saturation,brightness and contrast)and random rotation were used for data enhancement operation,and a total of 14147 strawberry datasets were obtained.The U-Net segmentation model was used as the prototype to improve the design.The encoder was modified based on the VGG16 model structure.The convolution with the Same was used to avoid the edge trimming operation before the feature map was merged,so as to retain more feature information and increase batch normalization which could speed up the training process and solve the problem of vanishing gradients.In this way,the experimental training speed and model training accuracy were improved,and the improved UNet model was evaluated by comparing with the original U-Net,PSPNet,and DeeplabV3 models under the same experimental data and parameters.[Results]The experimental results indicated that the method proposed in this paper had a better segmentation effect and better robustness on the selection of two types of test data sets:non-adhesive and non-adhesive.The recognition accuracy reached up to 96.05%,and the MIOU value was 89.41,while the F1_score value was 0.9.[Conclusion]The improved U-Net model displayed the best segmentation effects on the greenhouse strawberry dataset,and the results can provide strong technical support for subsequent strawberry automatic picking and quality grade identification.