Application Of Data Analysis Technology In Image Classification

With the large-scale infrastructure investment and the acceleration of industrialization in China,the output and consumption of the whole aluminum industry have increased rapidly,and China has become the world's largest aluminum production base and consumer market.After nearly 10 years of rapid growth,China's aluminum industry has entered a new stage of development,and shows many new trends.At present,China's aluminum processing industry is still in the state of "large size but not robust,or small scale but not proficient",forcing some high-precision aluminum required by the national economy and modern science still need to rely on foreign imports.In order to promote the development of the aluminum industry,on the one hand we need to be deep in the industrial field,on the other hand we need to strictly control the quality of aluminum production.The process of aluminum profile quality monitoring in industry is complex and needs a lot of professional theoretical knowledge,this paper will focus on how to solve the problem of aluminum profile quality testing through data analysis method.Quality inspection in this paper refers to the classification of 12 common defects on the surface of aluminum profiles.The traditional detection of surface defects of aluminum profiles is usually completed by sampling detection by experienced process engineers.This process is based on the sampling theory in statistics and the experience technology of process engineers,and has been highly praised by the industry.Compared with the traditional method of quality detection,using algorithm to detect quality can obtain higher accuracy at lower cost.As far as the task of this paper is concerned,there are two major challenges for the algorithm to perform quality detection.The first one is the serious imbalance between the number of defect categories,and the phenomenon of sample imbalance will lead to insufficient training of the model.The second is the over-fitting problem of the model.For some normal aluminum profiles with slight scratches,if the algorithm is too sensitive,the normal sample will be misjudged as a type of defect sample,or a type of defect sample will be misjudged as a normal sample with slight defects.Traditionally,scholars use statistical methods,signal processing methods and model-based methods to extract texture features of images,and then input feature vectors into classification models or neural networks for training.With the rapid development of computer technology,deep learning technology has been widely used,scholars are more likely to use deep learning framework to study the task of-image classification.Deep learning tends to perform better than traditional methods in classification accuracy,F1-score,confusion matrix,AUC and other evaluation indexes.First of all,in view of the serious imbalance between different types of aluminum profile data,this paper adopts content-based data enhancement and quantity-based data enhancement for data preprocessing before modeling.Then,considering the traditional method and the deep learning method,this paper designs eight different modeling ideas to compare the defect classification of aluminum profiles.On the traditional image classification modeling idea-feature vector combined with classifier,this paper selects three common texture feature extraction methods(HOG,LBP and SURF)and one texture feature extraction method proposed in this paper to extract defect features of aluminum profile images.Then,the feature vector is passed into the classification model to construct the classifier.In terms of classifier selection,this paper selects two statistical classification methods-LDA and QDA,and two machine learning classifiers-SVM and XGBoost.On the other hand,this paper uses two modeling methods,LeNet and ResNet50,to build the deep learning classification model.Finally,this paper uses the multi-classification accuracy,multi-classification accuracy,multi-classification recall rate,multi-classification F1-score and multi-classification AUC value as evaluation indicators,and evaluates the classification performance of the model mainly based on the multi-classification accuracy and multi-classification F1-score.Among the 31 models established in this paper,whether the multi-classification F1 value or the multi-classification accuracy is used as the evaluation index,the performance of the first five models is similar,which are XGBoost-Melt,XGBoost_Surf,QDA_Melt,LDA_Melt and SVM_Melt respectively.In the past,few scholars used XGBoost algorithm to solve the problem of image classification.However,the top two algor:ithms in this paper are all based on XGBoost method,which indicates that if the feature vector has strong representation ability on image,XGBoost algorithm can also achieve excellent classification performance on image classification.In addition,among the top 5 algorithms,4 methods all based on the Melt texture feature extraction method proposed in this paper,among which the top-ranked algorithm,XGBoost_Melt,has achieved over 90%multi-classification accuracy in the test set.To sum up,this paper innovatively combined XGBoost with HOG,LBP and SURF features respectively and achieved good classification effects.Meanwhile,the fusion feature Melt based on HOG,LBP and SURF proposed in this paper achieved good classification effects on different classifiers,which also provided a reference for other classification tasks.In addition,statistics,machine learning and deep learning methods were applied to improve the traditional aluminum profile quality detection scene,and more than 90%of the multi-classification accuracy was achieved,reflecting practical innovation.Industrial scenes are very different from each other,but as long as we do a specific analysis of specific problems according to industrial objects,we can extend the algorithm migration used in this paper to other industrial classification scenes.