Lightweight Image Classification System Based On Binary Discrete Cosine Transform

With the rapid development of information technology and automation technology in our society,researchers pay more and more attention to computer vision.Image classification system emerges as the times require.Image classification system is deployed on computer or computing platform to extract low-level image data into highlevel semantic features and realize image classification function.There are two standards to measure the performance of the image classification system.The first standard is the accuracy and the systems with higher accuracy are more reliable.The second standard is the algorithm complexity of the system,and the systems with low algorithm complexity are easier to deploy and run faster.The image classification systems based on convolutional neural network can extract image features adaptively.Compared with the image classification systems based on hand crafted features,the systems based on convolutional neural network have higher accuracy,but they consume more computing resources and storage resources.In this paper,a lightweight image classification system is designed based on binary discrete cosine transform algorithm.This paper completed the following work independently:(1)Design an image classification system.In the image feature extraction stage,the discrete cosine transform is used to extract features of the local blocks in the image effectively and quickly.In the feature reduction stage,the binary hash transform of multi graph merging and the block histogram transform are used to merge the features.Super parameter is introduced to simplify the extracted features.(2)The binary discrete cosine transform algorithm is used in image feature extraction stage.The two-dimensional discrete cosine transform is decomposed into orthogonal one-dimensional transform to reduce the algorithm complexity.The matrix DCT algorithm is combined with the binary DCT algorithm and the floating-point operation in the feature extraction process is replaced by the fixed-point operation,which greatly reduces the algorithm complexity.(3)The model is validated on two datasets:MNIST data set and VLOGO data set.Compared with the deep learning models and lightweight models,our image classification system has higher classification accuracy and lower algorithm complexity,which is easy to deploy in practical applications.