Low Illumination Image Enhancement Method Based On Non-separable Additive Wavelet And Filter Bank Optimization

Low illumination image is the image collected by optical imaging equipment at night or under weak illumination.This kind of image often has the problems of insufficient exposure,low contrast,blurred details and so on.A lot of detail information is submerged in the dark,which seriously affects users' observation and computer processing.The goal of low illumination image enhancement is to improve the brightness of the image,suppress the noise and correct the color deviation of the image to make the processed image more suitable for human visual observation or computer recognition.At present,the theory of two-dimensional non-separable wavelet has developed rapidly,and has achieved good application in image fusion,image segmentation,image enhancement and other fields.According to the characteristics of low illumination image and the advantages of two-dimensional non-separable wavelet,this paper proposes a low illumination image enhancement method based on two-dimensional non-separable additive wavelet.Since the two-dimensional non-separable wavelet is not restricted by the tensor product,the types and number of filter banks are infinite according to the given parameters.A large number of filter bank are uneven,so it is difficult to achieve the ideal processing result only by experience when selecting a filter bank.How to choose a better two-dimensional non-separable wavelet filter bank in the image enhancement process is still unsolved problem.With the rapid development of artificial intelligence and deep learning,deep learning models have been successfully applied in various fields of image processing.As one of the most representative deep learning models,CNN which is simple,convenient,efficient and accurate has been favored by majority of researchers.Therefore,this paper uses CNN model as a tool to research the selection and optimization of two-dimensional non-separable wavelet filter bank in low illumination image enhancement.The main innovations and research work of this paper are as follows:(1)This paper studies the theory of two-dimensional non-separable wavelet and additive wavelet,proposes a low illumination image enhancement method based on two-dimensional non-separable additive wavelet.Firstly,the image is decomposed into the illumination and reflectance.Then,three illumination with different characteristics are obtained by further enhancing of illumination.The three illuminations are decomposed by two-dimensional nonseparable additive wavelet,the low-frequency part is adopted the weighted average and the high-frequency part are used the fusion rule that the absolute value of the same position of the pixel is larger.Finally,the enhanced image is obtained by reconstruction.Experimental results show that the proposed method can effectively improve the brightness and contrast of the image,preserve the details of source image.It has a better performance in the objective and subjective quality evaluation than other six methods too.(2)Using CNN,the optimization and selection method of two-dimensional non-separable wavelet filter banks in low illumination image enhancement is proposed.Firstly,a large number of two-dimensional four channel non-separable wavelet filter banks are constructed,these filter banks are used for image enhancement by the method of(1),the enhancement results is quantified as the label of the filter banks to obtain the data set.Then the designed CNN network model is trained and tested by train set and test set.Finally,a small number of filter banks that do not belong to the original data set are reconstructed,the difference of the classification results and the real enhancement effect is compared.The experimental results show that the network has high accuracy in the test set,and the filter banks outside the data set are also identified accurately,which proves the effectiveness of the method.It has important guidance for the selection of two-dimensional non-separable wavelet filter banks in low illumination image enhancement.