| Image super-resolution technology is a popular research topic in the field of computer vision.Its advantages are low cost and good reconstruction effect.It has been widely used in medical images,astronomical images,surveillance images and other fields.In recent years,deep convolutional neural networks have performed well on image super-resolution tasks,but there are shortcomings such as large computation and slow real-time reconstruction;in addition,existing super-resolution algorithms pay little attention to feature channels Correlation between.Even if the channel attention module is introduced to increase the correlation between feature channels,the module is not applicable.In order to improve the shortcomings of the existing super-resolution,this paper proposes two improved algorithms:(1)Lightweight algorithm based on local enhancement(SR?fast).The SR?fast algorithm extracts deep-level features by cascading several enhancement blocks.Inside the enhancement block,channel split for feature maps,the relevant features are retained,and more useful features are extracted through multiple convolution operations,and finally all the features are fused.In terms of reducing the amount of parameters,compression modules and enhancement blocks that focus on local enhancement are mainly used for improvement.(2)Based on attention and local feature enhancement algorithm(SR?quality).The SR?quality algorithm is an improved version of SR?fast.The algorithm integrates the output of all enhancement blocks and adds a channel attention module of mixed pixel mean square error,global average pooling,and residual structure to obtain higher reconstruction quality.Experimental results show that,compared with the existing algorithms,the SR?fast algorithm proposed in this paper can guarantee the reconstruction quality while having the best real-time reconstruction speed and less parameter amount.The SR?quality algorithm has excellent reconstruction quality and detailed texture recovery. |
PDF Full Text Request