Research On Image Set Compression Technology For Deep Learning Classifier

With the rapid development of 5G technology in recent years,the mobile Internet has matured,and related applications of network applications and mobile terminals are emerging one after another.A large number of new applications are generated,which causes terminal devices to spend more storage capacity to store these applications and data.In daily life,images are the most frequently utilized by people,where more space is often needed to store them.Therefore,it is necessary to jointly compress many images.The quality and storage capacity of JPEG images are affected by many factors.In this paper,two more commonly used parameters,quality factor Q and image scale S are selected.By controlling these two parameters,the image is compressed.In this paper,the effect of two parameters on image storage capacity is deeply studied.Based on the structural similarity(SSIM)image quality measurement method,this paper proposes a compression method for image quality preservation.Under the condition that the image quality meets the actual demand,people hope to maximize the compression of images.In recent years,deep learning has achieved rapid development in practical applications,and image recognition and classification have been extensively studied.Therefore,this paper studies the image set double-parameter compression method for image categorization.Based on the quality factor and image scale of JPEG images,the classification accuracy is studied under convolutional neural networks(CNN)classifier.An image set compression method with classification accuracy is proposed.The experimental results show that compared with the traditional compression method,the proposed method can more fully compress many images and improve the classification accuracy.Further,this paper studies the adaptive compression method for the classification accuracy maintenance of online image sets.The proposed method still uses two compression parameters with quality factor and image scale.The background of the image captured by the image acquisition device in the same scene has certain similarities,and the images collected under 10 different scenes are combined into one image set with 10 types of images.The proposed method utilizes the image training set to obtain the classification model,and divides the image test set into several segments.Based on the similarity of the images,the compression method of the next small test set except the first one will refer to that of the previous small test set.The experimental results show that compared with the traditional image classification method,the classification accuracy of the proposed method is 3.3% higher than that of the existing method.Correspondingly,the online image set still maintains a large compression ratio.It can be inferred from the experimental results that if the amount of data in the online image set is increased,the image set compression ratio of the proposed method will be greatly improved.