Research On Surface Scar Identification Of Apple Based On Convolutional Neural Network

The classification of apple surface scars is a special application of image classification.The practice is to classify apples into a specific category based on the specific shape and size of apple epidermis scars.The traditional method of machine learning is to classify the problem of image classification by manually designing the features.However,in the face of the more complex problem of apple surface scar classification,the manual design features have a certain degree of complexity.In recent years,the deep learning method represented by the convolutional neural network algorithm has achieved great success in the field of image classification.Therefore,this paper will study the specific classification of apple surface scars through convolutional neural networks.The main research work and related results of this paper are as follows:(1)The construction of the database is the basis of the classification of the entire apple surface scar.This article has improved the database of apple surface scars of the research group to make the database samples more practical and convenient to operate.The database sample was taken from the field and contained 2130 two-dimensional images of the apple surface scar database,which contained 380 decayed states,580 round spots,570 scarless states,and 500 scratches.(2)The choice of network model has an important impact on the overall apple surface scar recognition rate.This paper uses the mature convolutional neural network model: Lenet,AlexNet,and GoogleNET to perform four states of decay/round spot/scratch/no scar.The classification experiment obtained the best recognition rate of 70%,73%,and 78%,respectively.This article selects AlexNet as the network model of the apple surface scar classification based on its own hardware environment and the convergence of the test set accuracy.(3)Based on the apple surface scar data and the AlexNet network model,the actual significance of the surface of the apple in this text was studied.The edible/inedible category,the scar/no scar classification,the round/strip/no scar classification The best recognition rates were 96.5%,99.6%,and 83.5%,respectively.Through the analysis of the experimental results,it is concluded that there are certain similarities between the round spot and the rotting state.(4)On the basis of convolutional neural network based on the apple surface scar classification experiment,aiming at the similarity problem,it proposes to use cluster segmentation and PCA dimensionality reduction to solve,and to be edible/data-enhanced data set.Inability to classify,with scar/no scar classification,round/strip/no scar classification,round spot/scratch/rot/no scarclassification experiment,yielded 96.5%,99.7%,83.5%,77% of the most The good recognition rate,to a certain extent,solved the similarity between the round spot sample and the rotten sample,and improved the recognition rate.