Using Convolutional Neural Network and Random Forest to Realize Potato Skin Defect Detection

Jialong Su; Zhenkai Wan

Volume : IX, Issue : VIII, August - 2020

Using Convolutional Neural Network and Random Forest to Realize Potato Skin Defect Detection

Jialong Su, Zhenkai Wan

Abstract :

Aiming at the problem that the potato skin defect detection algorithm does not fully extract the high-level features of the image and the detection accuracy is not high, a potato skin defect detection algorithm (CNN-RF) based on the combination of Convolution Neural Networks (CNN) and Random Forest (RF) models is proposed. In this paper, the final data set obtained by mirroring and random cropping methods from the images obtained by the laboratory machine vision platform was input into the convolutional neural network, and four types of defects, green skin, cracked groove, dry rot and budding data were used as the detection objects. Extract the high-level features of the images in the data set, then replace the Softmax layer of the convolutional neural network with a random forest, and use the trained output features as the input of the random forest model. In addition, use Gini impurity as the criterion for feature selection in the decision tree to verify the accuracy of the model�s recognition and detection of potato skin defects. The results show that the accuracy rate of the proposed CNN-RF is 97.4%, and its performance is better than the detection effect of the Particle Swarm Optimization- support vector machines(PSO-SVM), AdaBoost, hyperspectral data dimensionality reduction method, Back Propagation Neural Network (BPNN) and conventional CNN model.