Research On Convolutional Neural Network Architecture Search

Convolutional Neural Network(CNN)has made great achievements in the field of image feature learning,and has become a mainstream method to solove image classification tasks,image denoising tasks,object detection tasks,image segmentation tasks,and other challenging computer vision tasks.From the overlay of network layers to the residual architectures and densely connected architectures,and then to the Google Net,network architecture has obviously become the key issue to image feature learning ability of CNNs.However,traditional manual design of CNN architectures is not easy.Not only lots of CNN-related professional knowledge and experience are required,but also the manual design of different CNN network structures for different datasets is inefficient.These problems limit the representation learning ability of CNNs.To alleviate the difficulty of manual design,automatic architecture design algorithms have been proposed recently.However,automatic architecture design algorithms usually consume considerable computational time and resources.And only deep blocks have been appointed in designing without considering the wide blocks of CNNs,which limits the performance of the evolved CNNs.In addition,existing NAS algorithms are proposed to solve image classification tasks.However,there are also problems existing in designing the CNN architectures which are efficient to solve image denoising tasks.NAS algorithms for image denoising tasks are still in their infancy.In order to slove the above problems in existing automatic CNN,an improved algorithm is proposed in this paper based on the related research on the CNN network architectures and NAS algorithms and the following work has been carried out:(1)To solve the time-consuming and resource-dependent problem existing in the above automatic CNN design algorithms,a refining fitness evaluation method is proposed to improve the reliability of the partial-datasetsbased automatic architecture design algorithms and reduce the running time.A regression model is trained based on CNN model parameters,the accuracy of the CNN architecture model in the divided dataset,and the accuracy of the CNN architecture model in the complete dataset.At the same time,an adaptive regularization penalty mechanism is introduced to further ensure the effectiveness of the above regression model.To slove the problem that proposed algorithms only use deep architectures and ignore wide architectures of the CNN,the Inception Block is introduced introduced in this paper to evolve flexible and optimal CNN architectures from both deep and wide architectures.In addition,the Feature Block,the Transition Block,and the Dropout Block are proposed for feature extraction of shallow network layers,convolutional and pooling operation,and prevention of overfitting,respectively.Experimental results on image classification datasets CIFAR-10 and CIFAR-100 show that the proposed algorithm can automatically design CNN architectures with competitive feature learning ability in very short time under the premise of reliability and limited computing resources.(2)In order to solve the problem that proposed NAS algorithms are not effective for image denoising tasks,a new NAS algorithm based on DnCNN for automatic image denoising CNN architecture design is proposed in this paper.As for the simple overlay architectures of DnCNN algorithm,the deep CNN architectures,Res Net Block and Dense Net Block,and the wide CNN architecture,Inception Block,are introduced in the proposed algorithms.In addition,in consideration of the characteristics of the image denoising tasks,the Feature Block,the Transition Block,and the fitness function are modified crospondingly.Experiments results on image denoising datasets S12 and BSD68 shows that the proposed algorithm can automatically design competitive image denoising CNN architectures quickly.