Research on the Surface Damage Detection of Gastrodia elata Blume Based on Convolutional Neural Network

PurposeIn order to solve the problem that the traditional Gastrodia elata Blume surface damage mainly depends on manual inspection, we proposed to apply the residual neural network model (Faster R-CNN ResNet101) to the identification of G. elata surface damage, in order to obtain a better identification effect.

MethodFour kinds of G. elata were studied, including decay, mildew, mechanical damage and intact G. elata. Firstly, the model was constructed on the basis of convolutional neural network and regional candidate network, and the model was tested on tensorflow framework, finally, the results were compared and analyzed.

ResultThe surface damage detection model of G. elata was based on the input convolution layer and four convolution groups in Fast R-CNN ResNet101 network, and the region candidate network was used to generate the initial position candidate box of the surface damage detection, to classify and locate the candidate boxes, the recognition rate was 95.14%, the precision rate was 0.94 and the recall rate was 0.92. Compared with SSD (Single Shot multibox Detector), Faster_rcnn_inception and Rfcn_resnet101, the accuracy of network recognition was improved by 13.02%, 10.69% and 12.02%, respectively.

ConclusionThe model has the characteristics of strong generalization ability, high accuracy and good stick property, which can be used as a reference for the identification of agricultural products.