Research On ELM Image Classification Combining HOG And Random Forest

In this era,the internet industry is booming.People can upload any images at anytime and anywhere.With the popularity of intelligent machines and the promotion ofsocial platforms,image sharing has become the current trend.Thus,it is obvious that image data will be an important part of the whole internet information.How to understanding the content of these images is a crucial problem for find the user behavior patterns and mining new knowledge.Image classification is animportant method to understand the content of the image,and has been successfully applied in the fields of finance,postal service,public security and transportation.The key point of image classification are feature extraction and classifier.Feature extraction can extract the key information from original image,and contribute to improve accuracy for image classification.The image classification methods based on image feature extraction have many successful applications.The classifier is able to construct the map function from input to output.The classifier can optimize the classification results by a large number of data.But the problem is: a.For feature extraction,artificial select the feature extraction is uncertainly.And the extracted features are not optimized.To the certain extent,a part of the extracted featureshave negative influence on accuracy of classification;b.For classification,to get a better classification accuracy usually need long training time.How to balance the relationship between training time and accuracy is a considerable problem.The main innovation of this paper is to propose a novel Extreme Learning Machine(ELM)classification model based on Histogram of Oriented Gradient(HOG)and Random Forest(RF).By extracting HOG feature of the image,and introduce Random Forest method to measure the importance of HOG vector.Finally,the important vector is selected as input of Extreme Learning Machine.We performed the proposed method on MNIST and USPS datasets.The experimentresults show that the proposed methodhas higher accuracy than HOG-ELM?ELM and ML-ELM.And has faster trainingspeed than ML-ELM.