Fruit Classification And Implementation Based On Sparse Representation And Convolutional Neural Network

The classification technology of fruit has been a hot issue in the field of agriculture information at home and abroad,and has made great progress in recent years.But the recognition rate of fruit is significantly decreased under the condition of large scale,complex data,partial occlusion,illumination changes and complex scene.As a result,the application of fruit recognition technology to the actual agriculture still faces many problems and challenges.Aiming at the problem of large scale fruit species identification,redundant features,manual design feature is subjective problems.From the fruit characteristics of the expression of feature technology and automatic learning two aspects of research: In the aspect of feature expression,the sparse coding method of RGB fruit image and the local constraint linear coding method of RGB fruit image are studied;in the feature automatic learning,fruit classification method based on convolutional neural network is proposed.The main work of this paper is as follows:(1)Proposed matching model based on sparse encoding space in Pyramid(Sparse Coding Spatial Pyramid Matching,Sc SPM)fruit classification method.Firstly,the SIFT local feature points were extracted in RGB color fruit image,to overcome the effects of the different illumination and scale transform.Then the sparse coding and spatial pyramid feature expression.Finally,the linear SVM classifiers achieve fruit species identification.The experimental results show that the classification accuracy rate of fruit is 98.90%,and the accuracy of the method is 8.02% higher than that of the color histogram feature extraction after the fusion of SIFT feature and Sc SPM feature expression.(2)Locality-constrained Coding(LLC)fruit classification method based on local constraint linear coding in spatial Pyramid matching(Linear)model is proposed.Because of the space matching model in Pyramid sparse encoding image encoding speed is relatively slow,the local linear constraints for encoding spatial Pyramid matching model to improve the speed of image encoding.LLC considering the sparsity and local characteristics,types of fruit classification accuracy rate was 99.25%.(3)Fruit classification methods based on convolutional neural networks(Neural Networks Convolutional,CNN)are proposed.The method uses a deep learning framework--Caffe platform,using the CNN of the image are essential characteristics of automatic learning,fruits are correct classification rate reached 99.53%,overcomes the manual feature design depends on the experience,to improve the accuracy of the classification method of fruits.(4)On the basis of the above research,the development and implementation of the fruit classification system was carried out.Using camera sensor in illumination changes,360 degree perspective transformation and partial occlusion environment under collected apple,passion fruit,dragon fruit,such as oranges,14 kinds of 39 kinds of different types of fruit image 96815 amplitude,the establishment of large-scale fruit RGB image database.Fruit classification system is designed,using MATLAB image processing tools to develop a system interface,to the fruit classification algorithm programming try the three RGB fruit classification algorithm are analyzed and compared.In fruit classification experiments,the three kinds of algorithm are high accuracy,which based on the CNN method accuracy is the highest,the LLC algorithm times,Sc SPM algorithm minimum;in time performance,Sc SPM algorithm consume time the longest,LLC algorithm secondly,CNN algorithm transport time shortest.Compared with common color histogram feature extraction methods,CNN,LLC and Sc SPM feature extraction method is 8.65%,8.37% and 8.02%,respectively,which is higher than the color histogram feature extraction method.Based on the above experimental results shows that fruit classification,CNN algorithm compared with Sc SPM algorithm and LLC algorithm with linear encoding constraint classification performance better.