Research On The Detection Method Of Potato Buds Based On Convolutional Neural Networks

As one of the main food crops in countries around the world,potato plays an important role in maintaining world food safety and stability,especially in developing countries.At present,the cutting of seed potatoes is mainly done manually due to its highly required professional skills.But the seed cuts completed by different workers have significant differences.The challenges of manual cutting for seed potatoes is increasingly severe with the problem of high labor intensity and high cost,especially with the increasing labor costs and the reducing labors.The problem of the automated cutting of seed potatoes needs to be urgently solved.Potato bud detection is a key factor to realize the automated cutting of seed potatoes,and the detection performance directly affects the later quality of seed cuts,and thus affects the potato yield.At present,researches on the potato bud detection are limited.The existing methods about potato bud detection and segmentation are based on the traditional approach,which is in poor detection performance as well as time-consuming and can not lay a solid foundation for the automated cutting of seed potatoes.In order to improve the detection time and detection performance,thus guaranteeing the quality of seed cuts and the potato yield,we propose a detection method for potato buds based on convolutional neural networks.The main research contents and innovations are as follows:(1)The detection performance and detection time of different convolutional neural netwoks in potato bud detection are studied with the approach of transfer learning and“fine-tuning”.With the Faster R-CNN framework,transfer learning technology is carried out on eight common convolutional neural networks including AlexNet,VGG-16,VGG-19,GoogleNet,SqueezeNet,ResNet-34,ResNet-50 and ResNet-101 which have been pre-trained on the ImageNet,and“fine-tuning”operation is performed according to the structure of Faster R-CNN framework.ResNet-50 is chosen as the backbone network of Faster R-CNN framework by comparing the detection performance and detection time of different convolutional neural networks.On this basis,a series of optimized strategies are proposed to improve the detection performance of potato buds.(2)A method of multi-scale feature connection is proposed in the research.In this method,feature maps generated by the Res2c layer of the second convolution layer and the Res3d layer of the third convolution layer are taken as the input of the ROI pooling along with the feature map generated by the fourth convolution layer,so that the Faster R-CNN model can learn more features about the object and improve the classification ability of the model.Besides,a 1×1 convolution layer is used for dimensionality reduction and control the size of the model.(3)An optimized non-maximum suppression algorithm is proposed in the research.In the optimized algorithm,for the detection boxes whose IOU with the detection box with the highest score are greater than or equal to N_t,the Gaussian function with IOU is leveraged to decay the confidence of them.Then the decayed confidence is compared with the discriminant parameter O_t.When the decayed confidence is less than the discriminant parameter O_t,the current detection box is removed;when the decayed confidence is greater than the discriminant parameter O_t,the current detection box is retained to improve the detection performance of the model.(4)An improved method of the default anchors based on the chaos optimization K-means algorithm is proposed in the research.To prevent K-means algorithm falling into a local minimum,the variable in K-means algorithm is optimized by the chaotic variable.First,map the Logistic chaotic variable into the range of the variable in K-means algorithm,and perform a global search with the chaotic variable at a higher speed.Second,the optimized variable is leveraged to generate the anchors which are more suitable to the size of potato buds by performing cluster analysis on the labels of potato buds in the training set,thus improving the localization ability of the model.Experimental results reveal that the detection precision and recall rate of the improved Faster R-CNN model are 3.98%and 9.01%higher than those of the original Faster R-CNN.Moreover,the average detection time of the improved Faster R-CNN is 0.166s,which is the same as that of the original Faster R-CNN.Namely,the improved Faster R-CNN model can boost the detection performance of potato buds without occurring any noticeable computational overhead.Consequently,the improved Faster R-CNN model satisfies the requirements for real-time processing and can lay a solid foundation for the automated cutting of seed potatoes.