Research On Infrared Image Restoration Based On Convolutional Neural Networks

The important issues of infrared image restoration include the low resolution and significant non-uniform noise.Differ from traditional algorithms,the method of infrared image restoration based on convolutional neural networks can make the algorithm more intelligent in improving restoration ability and reducing design cost.However,these methods only can deal with single-task processing(i.e.,denoising or super-resolution).Meanwhile,the need for a rich amount of data and the training mode of large-scale network structure causes the difficulty of data collection,model training,and lack of storage.In response to these problems and difficulties,this article focuses on the following research work:(1)We propose a model cascading two convolutional neural networks to achieve feature reconstruction step by step,which deals with the difficulty of extracting detailed features in the infrared image super-resolution under the condition of small sample training.Firstly,we use the convolutional and transpose convolutional sub-network to extract the edge features quickly.Then,the recursive residual sub-network is used to further extract the texture features.Finally,our model restored by the global residual learning.The experimental results show that the proposed model has the comparable ability as the comparison algorithm under the condition of fewer parameters.(2)So as to the over-fitting problem by too many network structure parameters and the need for multi-task infrared image restoration under small-scale training,we propose a restoration method using multi-cascade group convolutional blocks,which achieve Gaussian denoising and super-resolution reconstruction at the same time under fewer conditions of training data and times.Firstly,multi-scale feature extraction is set up by different sizes of convolution kernels to reduce the demand for data.Secondly,the depthwise convolution and pointwise convolution are used to alternately complete the denoising and reconstruction in the module and the convolution calculation amount is compressed.In this way,the block-to-block cascade deepens high-frequency feature learning,and local residuals ensure fast convergence.Finally,infrared images are restored by global residual feature fusion.We select a better model from the kernel size and number in the stage of multi-scale feature extraction and the blocks internal structure and number in the stage of cascaded group convolution.The experiment reveals the impact of different strategies on the performance of the model from the dataset setting,optimization algorithm,and network sparsity training.Meanwhile,we compare the proposed method with the state-of-the-art algorithms and analyze the effect of batch normalization layer and different residual learning forms.Experimental results show that the proposed model with fewer parameters than other models can complete high-quality infrared image restoration with fewer data and training times.At the same time,it has obvious advantages in comparison of generalization performance indicators under different noise variances and reconstruction multiples.