| With the continuous development and innovation of technology,amazing breakthroughs has been made on deep learning theory and its applications,especially in the application of computer vision.The image super-resolution reconstruction technology has always been a hot issue in the field of computer vision.Super-resolution reconstruction technology refers to converting low-resolution images into higher-quality high-resolution images while maintaining the unique structural information of the image.Algorithms for image super resolution reconstruction can be divided into traditional reconstruction algorithms and reconstruction algorithms based on deep learning.The super-resolution reconstruction algorithms based on deep learning are more effective than the traditional methods.Although the super resolution reconstruction algorithm based on deep learning can achieve good results,there are some shortages of it,such as high computation complication,large number of parameters and huge redundancy,which is not conducive to the implementation in mobile devices.In order to obtain a portable super resolution reconstruction algorithm,this paper is based on deep learning technology,combines GAN network,the quantization of CNN's weight,network structure adjustment and other methods to optimize the super-resolution reconstruction algorithm,and designs a super resolution reconstruction APP based on the Android platform.The content including:(1)The super resolution reconstruction technology algorithms are studied in detail on a general algorithm and its performance.The theory and technique of deep learning in recent years is discussed,especially on the super resolution reconstruction algorithms based on deep learning.A new super resolution reconstruction algorithm in this thesis is proposed according to these previous ones.(2)In order to obtain a large number of training sets,a large number of the high definition on-board images from the car video recorder and the down-sampled images of these images are respectively used as two parts of the training dataset for the deep learning.In addition,cropping,rotation,and folding are used to further enhance the data.(3)The adversarial generation network is introduced as the key part of the super-resolution reconstruction model,In which the quantization of weight is used to makethe size of the algorithm small.Firstly,a quantized convolution weight module is implemented based on the weighted method,and the information compression module is designed by combining the characteristics of the residual structure and the dense network to reduce the redundancy among the parameters and strengthen the relevance between the weights.Then,the main part of the network is composed of a feature extraction module,multiple stacked information compression modules,and a reconstruction module,and the global skip connection is used to fuse the information in the super-resolution space after the bicubic interpolation with the image after the low-resolution space reconstruction.On the basis of not increasing the complexity of the model,improve the ability of network expression.Finally,the network is nested into the adversarial generative network,and the network is trained by the dataset.(4)In order to apply the proposed model to a mobile device,the corresponding Android APP is developed.This APP takes up less memory than other deep learning models by the quantification of the parameters,which is suitable to the mobile intelligent terminals.The super-resolution reconstruction algorithm based on deep learning designed in this paper can get good subjective visual experience,and it is lightweight and efficient.Compared with the Lap SRN algorithm,the model in this paper takes up less than 2/5 of its memory,and the time required for reconstruction is less than 1/7.The APP designed based on the algorithm in this paper can still obtain high-definition high-definition images under the condition of limited hardware equipment. |
PDF Full Text Request