Research On Image Classification Method Based On Deep Second-order Statistical Features

Image classification technology is the most basic and important branch of computer vision.It has a wide application market in face recognition,intelligent health care and mobile payment.Among them,texture image classification is a very important technology in image classification.It has huge application potential in the fields of material texture recognition and pipeline detection.With the rapid development of a series of social media applications and short video applications,the large-scale increase in the number of images,it is becoming increasingly important to obtain image features with more discriminative capabilities on large-scale images.Because the deep convolutional neural network has strong feature learning capabilities,it can generate image expressions with strong discriminative ability,and deep second-order statistical features can be obtained from the deep feature itself,which has higher computing efficiency and can be more robust image expression.Therefore,this paper explores how to effectively combine the second-order statistical feature information with depth features and the attention mechanism.The main work can be summarized as follows:(1)A Robust Deep Gaussian Descriptor(RDGD)based on bilinear convolution features is proposed and effectively applied to image tasks such as texture datasets.RDGD combines a bilinear convolutional neural network(B-CNN)and a Gaussian descriptor as a new texture representation method.The outer product calculated by the B-CNN is embedded into the Gaussian expression as a rough estimate of the covariance.At the same time,in order to overcome the case of high-dimensional small samples,the estimated sample covariance is not robust.Based on the previous step,a matrix power normalization operation is used to eliminate the effect of rough estimation of covariance.RDGD is evaluated on three widely used texture datasets.The experimental results show that RDGD is better than its benchmark network B-CNN,and it is also comparable to the state-of-the-arts at that time.(2)A Second-order Response Transform Attention Network(SoRTA-Net)is proposed,which is an image classification network framework based on the combination of the second-order response conversion mechanism and the attention mechanism under computer vision.By calculating the second-order statistical characteristics of the depth features and applying the attention mechanism on this basis,the ability of deep convolutional neural networks to express images is greatly enhanced.In order to explore a more flexible second-order response conversion mechanism,a Refined Second-Order Response Transform(RSoRT)module(a network formed by the combination of these modules is called SoRTA-Net)is proposed,which can specifically correct the characteristic response.At the same time,it is possible to robustly model second-order features of complex image features.Through extensive evaluation on a series of data sets,the results show that its performance is significantly improved compared to its benchmark method SoRT,and SoRTA-Net is also comparable to the best method at the same time.