Research On Rotation Invariant Image Feature Representation Based On Deep Learning

Rotation invariant feature representation plays important roles in a variety of computer vision tasks such as key-point matching,object detection,image retrieval,and image classification.Due to the powerful ability of convolutional neural networks(CNNs)to learn complex feature representations,the exploitation of invariant information based on CNNs has become an important topic in the field of computer vision.However,traditional CNNs have limited ability to learn rotation invariant features due to the lack of inherent mechanisms to handle rotation changes.To this end,this thesis carries out the following research work:1.A rotation invariant CNN(RICNN)based on orientation pooling and covariance pooling is proposed to learn rotation invariant deep higher-order features.First,RICNN employs the rotating convolution operator to obtain multi-directional filter responses.Then,RICNN performs two types of orientation pooling(OP),i.e.,maximum OP and average OP,thereby achieving pixel-level rotation invariant feature representations.Meanwhile,RICNN embeds covariance pooling(CP)to learn second-order rotation invariant features,thereby achieving image-level rotation invariant second-order feature representations.Finally,the above features are concatenated and used for image classification with the cross-entropy loss.Experiments show that RICNN is invariant to image rotation and obtains excellent classification performance.2.A rotation invariant Gabor CNN(RIGCN)is proposed to learn rotation invariant deep features by using Gabor prior information.First,a transformation is applied to each input image to generate multiple rotated image instances.Then,these image instances are fed into weight-sharing Siamese networks to learn Gabor-guided deep convolutional features.Next,the maximum and average feature responses are computed from all the rotated instances of the same input image and fed into a 1D convolution based fusion module to obtain a rotation invariant image representation.Finally,the cross-entropy loss is used for image classification.In this method,the use of Siamese architecture enables RIGCN to directly learn rotation invariant features from rotated image instances.The use of convolutional fusion module enables RIGCN to obtain rich statistical features with high efficiency.Experimental results demonstrate the effectiveness of RIGCN for rotation invariant image classification.