Architecture Design And Application Of Deep Neural Network

With the advent of big data era and the continuous improvement of computing capabilities,deep learning has made many breakthroughs in the fields of image classification,speech recognition,natural language processing,etc.,which has greatly promoted the vigorous development of artificial intelligence.Deep learning is supported by big data,and its performance is greatly improved at the expense of extremely high computational complexity and storage space,however,the promotion and application of deep learning are limited by computational complexity and storage space.For specific tasks and application requirements,how to design specific deep neural network architectures,which greatly reduce the computational complexity without affecting performance,is currently a hot topic in the field of deep learning.Convolutional Neural Network(CNN)is one of the most widely used deep learning methods in computer vision and image processing.For specific applications such as extreme weather recognition,pulmonary nodule malignant lesion recognition,and multi-degraded image restoration,the design of CNN architectures is studied deeply in this paper,which reduces complexity of network without sacrificing performance.The contributions in this paper provide a good foundation for the promotion and application of CNN.The research content of this paper includes the following three aspects:(1)An extreme weather recognition method based on dual fine-tuning strategy and truncated GoogLeNet is proposed in this paper.Since weather is affected by many factors and the differences between extreme weather images are small,the traditional manual design features are difficult to accurately represent various complex weather image characteristics,resulting in low accuracy of extreme weather recognition.Firstly,for the specific application requirement of extreme weather recognition,a large-scale weather image dataset “WeatherDataset” is constructed in this paper,in which 16635 extreme weather images are divided into four classes(sunny,rain,snow,and fog),and complex scenes such as urban roads and highways are coverd.Next,various features of extreme weather images can be extracted by GoogLeNet from big data,and then extreme weather recognition model can be established.Finally,due to the complexity of GoogLeNet is relatively high,GoogLeNet is truncated to obtain a lightweight network architecture,and then a dual fine-tuning strategy is proposed to optimize the network parameters.The experimental results have demonstrated that,compared with the original GoogLeNet,the recognition accuracy of the proposed method increases from 94.74% to 95.46%.The model size of the proposed method is only 31.23% of original GoogLeNet.On CPU and GPU implementation,the proposed method processes the images 1.39 times faster and 2.44 faster than the original GoogLeNet respectively.(2)A pulmonary nodule malignant lesion recognition method based on lightweight CNN is proposed in this paper.Since the number of pulmonary nodule malignant lesion images is small and the use of deep CNN is prone to overfitting,a lightweight CNN is designed.The network includes only 6 parameter layers,which are 4 convolution layers and 2 fully connected layers.Compared with several typical CNN architectures such as AlexNet,GoogLeNet,etc.,the architecture of CNN designed in this paper has obvious advantages in the three aspects of recognition accuracy,recognition speed,and model size.The recognition accuracy can reach 96.42%.(3)A multi-degraded image restoration method based on CNN and multi-model fusion is proposed in this paper.Firstly,for degraded images affected by three kinds of mixed degrading factors such as gaussian noise,downsampling,and compression distortion,a CNN architecture is designed,in which the input is a degraded image and the output is a residual image.A mapping model of degraded-residual image can be obtained through training,which is used to restore degraded image.Next,the multi-model fusion is used to post-process the restored image.The image is divided into blocks,and the image blocks are divided into three categories according to the texture complexity of the image blocks,such as low texture complexity,medium texture complexity,and high texture complexity,then the corresponding three image restoration models are respectively learned.Finally,a restored image fusion method is proposed.The degraded image is respectively input into three restoration models to obtain three restored images,and these restored images are fused to obtain a final restored image.The experimental results have demonstrated that,compared with single restoration model,the multi-model fusion method proposed in this paper can effectively improve the performance of image restoration.