Feature Pooling:A Feature Selection Method Used In Convolutional Neural Network

Convolutional Neural Network(CNN)is one of the main research directions of deep learning and plays a crucial role in many fields,especially in the field of Computer Vision.CNN's powerful feature extraction and feature representation capabilities have made traditional feature engineering based solutions in the visual field increasingly obsolete.The research focus has gradually shifted to how to design CNN frameworks that are more effective for specific tasks.Looking at the representative CNN framework in recent years,one prominent change is that the network has become deeper and wider.Which also brought a series of problems in the network training stage,like network over-fitting,gradient vanish and network degradation problem.In order to solve the problems caused by deep neural networks,some solutions that can effectively alleviate such problem are proposed,such as Batch Normalization,Residual Learning,etc.However,there is still a bottleneck in increasing the depth of the network.Exceeding a certain depth range will lead to a smaller accuracy revenue,and may even be counterproductive.How to design or improve the existing CNN architecture and further enhance the generalization ability of the neural network is still a research topic with broad research prospects.In view of the above problems,we proposes the idea of feature selection within CNN.Through the feature selection process,some effective features are filtered,which reduces the risk of over-fitting and enhances the performance of the network to some extent.The main research contents has the following four aspects:1.Analyzing the design and improving the existing CNN architecture still has practical significance.Starting from the training problems,many solutions that have emerged are analyzed,and there is still a bottleneck in the depth increase.The existing CNN structure still has the possibility of improvement.2.Proposed the idea of explicit feature selection within CNN.Combined feature selection procedure with efficient GPU computing,we designed the Feature Pooling algorithm and gives the specific step,and using Tensor Flow as the implementation framework toimplement it.3.Verify the feasibility,effectiveness and versatility of Feature Pooling in CNN.Selected three typical visual tasks: image denoising,image classification and image style transfer.Designed and implemented the specific neural network,and the concrete analysis of the experimental results.4.In view of the characteristics of Feature Pooling,it compares the differences with related theories and analyzes its own advantages and disadvantages.At the same time,the concrete suggestions for using Feature Pooling in practical applications are given.