基于CNN的二维四通道不可分小波滤波器识别

2D 4⁃channel non⁃separable wavelet filter recognition based on CNN

二维四通道不可分小波在图像融合中得到了成功的应用,但其滤波器组的选取都是依靠经验手动完成的,这样会导致图像融合的结果难以达到最优。针对在图像融合的过程中选择融合效果最优的滤波器组的问题,提出基于CNN的二维四通道不可分小波滤波器组的识别方法。该方法利用卷积神经网络强大的学习功能,对构造的大量二维四通道不可分小波滤波器组进行训练。让训练好的网络对滤波器组进行识别,让这些识别的不同类别滤波器组分别作图像融合,并对比其融合效果。实验结果表明,该方法在测试集上的识别准确率达到0.999 5,对训练测试集外的滤波器组同样识别准确,解决了图像融合过程中二维不可分小波滤波器组选取困难的问题。

The 2D 4⁃channel non⁃separable wavelet has been successfully applied in the image fusion,but the selection of its filter bank is manually completed by relying on experience,which makes it difficult to achieve the optimal result of image fusion.How to choose a filter bank with the optimal fusion effect in the process of image fusion is still an outstanding issue.In view of the above,a CNN⁃based recognition method of 2D 4⁃channel non⁃separable wavelet filter banks is proposed.In this method,a large number of constructed 2D 4⁃channel non⁃separable wavelet filter banks are trained by virtue of the powerful learning function of CNN.The trained network is used to recognize the filter banks,and then these different types of filter banks identified are used for image fusion respectively and their fusion effects are contrasted.The experimental results show that the proposed method′s recognition accuracy on testing set is 0.9995.The method also has an accurate recognition for those filter banks outside the testing set.It has been verified that the method can solve the difficulty in selecting 2D non⁃separable wavelet filter banks in the process of image fusion.