The Research On Medicine Image Super Resolution Algorithm Based On Deep Neural Network

Due to the influence of hardware itself,technical level and various external environmental factors in the traditional imaging system,the image generated in reality will inevitably lose some original detailed texture information.In order to effectively solve such problems,image super-resolution technology has been extensively studied.This technology can restore and reconstruct images with poor quality,and has the advantages of low cost and high universality.For the application of image super-resolution technology in medical image reconstruction,although the current deep learning-based method can improve the image resolution to a certain extent,the problem of large calculation amount has not been effectively solved.Secondly,most models do not fully use the hierarchical features of the original image,lack discriminative learning ability across feature channels,and do not make effective use of global and local features.In order to alleviate this series of problems,this paper improves several existing super-point network models.The specific research contents are as follows:(1)An image super-resolution algorithm based on octave convolution dense sampling block network is proposed.First,the octave convolution is referenced in the initial feature extraction stage,and the octave convolution operation is used to process the information of different frequency distributions in the feature.The process of initial feature extraction is divided into high-frequency operation and low-frequency operation,where the main function of high-frequency operation is to restore as much high-frequency information as possible,and forward these features to high-resolution output,so that the main network process concentrates on learning high Frequency information,thereby reducing unnecessary calculations.Secondly,adding an improved dense jump connection sampling block to the network provides an effective method for the combination of low-frequency features and high-frequency features,and constructs a short path directly from the output to each layer,which can effectively alleviate the gradient disappearance of the deep network problem.At the same time,a powerful high-level representation is generated by the sampling block to improve the reconstruction function.Finally,the experimental results show that the improved algorithm has shown the superiority of the experiment in terms of calculation amount,objective evaluation indicators,and subjective visual experience compared to common algorithms on the medical image data set used,thus verifying the proposed method.Effectiveness.(2)An image super-resolution algorithm based on adaptive residual neural network is proposed.First,an adaptive framework is designed to switch the global and local inference in a flexible way to the internal features of the image,so that a large number of global features can be extracted without ignoring key information,which is conducive to the comprehensiveness of the residual image and thus recovers High-quality images.The Squeeze-and-Excitation Networks(SE-Net)structure is added to the designed adaptive module,and the extracted features are channel modeled.The learning method enables the network to autonomously obtain the importance of each feature channel.You can enhance the useful features in the image reconstruction process and suppress the useless or less useful features according to the obtained importance.In this way,the deep features extracted by the network have more nonlinearity,which can better fit the complex correlation in the feature channel,while effectively reducing the amount of parameters and calculations,and can improve the performance of super-resolution to a certain extent.Finally,according to the experimental results,the algorithm has a positive impact on the network layer and SE-Net research,which makes the objective evaluation index value higher than the commonly used algorithm,and the image generated by the algorithm can be seen more clearly,the details are more obvious.