Research And Application Of Rotation Invariance Of Convolutional Neural Network

With the development of deep learning,especially the development of Convolutional Neural Networks(CNN),image classification and object detection methods have made significant progress,but even the most advanced models have only very limited rotation invariance.When the input data is rotated in space,since the filter of the original model is polished and learned from the upright sample,the model regards the image features before and after the rotation as two completely different categories,resulting in a significant reduction in the final recognition effect.Known solutions include the enhancement of training data and the increase of rotation invariance by globally merging the rotation equivariant features.These methods either increase the workload of training or increase the number of model parameters.In order to solve the above problems,this paper introduces a module that can be inserted into the existing network,and directly incorporates the rotation invariance into the feature extraction part of the convolutional neural network.This module does not have learnable parameters and will not increase the complexity of the model.At the same time,only by training the unrotated data,it can perform well on the rotated test set.These advantages will be suitable for fields such as biomedicine and astronomy where it is difficult to obtain samples and the target has no directionality,as well as the fields that have higher requirements for imaging speed,such as industry.The main research content of this thesis includes the following four parts:1.In view of the lack of rotation invariance of convolutional neural networks,this paper makes the feature layers before and after convolution satisfy the rotation equivariance,and finally makes the results of the entire neural network have rotation invariance.After deduction,when the rotation over angles?{0,90,180,270}~°,the feature code is exactly the same.When at any rotation angle,there is a small distinction between feature encodes.2.The design is based on the Regional Rotation Layer(RRL)of the improved LBP operator,which is embedded in the convolutional neural network in the form of a plug-in,without the need for substantial changes to the context.The position of the neighboring pixels is related to the maximum LBP descriptor of the pattern;the number of times to implement improved LBP is determined by the convolution step size and the size of the feature map.When the convolution kernel slides on the feature layer with a sliding window,the area covered each time It is the location for the implementation of the improved LBP operator.3.Insert the Regional Rotation Layer into the existing convolutional neural network,including the common standard convolution structure and the residual structure used on a large scale,as well as the depth separable convolution.Choose the backbone network that performs well in classification tasks and target detection tasks,integrate the Regional Rotation Layer,and explore the fusion method that can play the most role.Inference training on a variety of data sets verifies the effectiveness of the improved model.4.Designed and implemented a pneumonia detection system.The core algorithm module uses a target detection network with a Regional Rotation Layer.The user submits an X-ray of the patient's lungs on the system to obtain the predicted area of the lesion to assist in the diagnosis.