面向残差网络多元特征的轻量级虹膜分类

针对传统虹膜分类需手工设计滤波器提取虹膜特征, 提取特征单一, 且通常需大量手工调参, 泛化能力较差的问题, 提出一种面向残差网络下多元特征的虹膜分类算法. 一方面将虹膜图像与Gabor特征相结合, 另一方面在网络结构中使用多个尺度的卷积核, 使学习到的虹膜特征更丰富, 从而提高图像特征的表征能力. 实验结果表明, 在固定类别中, 使用Softmax分类器进行多分类, 该算法在JLU虹膜数据库中的分类准确率可稳定在98.90%以上, 不低于DeepIrisNet和Resnet等网络结构, 且该算法的网络结构参数更少， 学习速度更快.

Linear Projection Decolorization Algorithm Based on Structural Information Similarity

Aiming at the problem that the structural information were easily lost in the process of decolorization, we proposed a decoloriz ation algorithm based on structural information similarity. Firstly, the contrast image was constructed in RGB color space by using the average and standard deviation of pixels to keep the contrast and brightness information of the color image. Secondly, the similarity between each RGB channel image and contrast image was measured by structural similarity index. Finally, the structural similarity indices were taken as the weights in the global weighted mapping function to obtain the final grayscale image. This algorithm effectively solved the problem that the objective function needed to be solved in other typical decolorization algorithms, which resulted in the defects of high complexity and unnatural structural information. The experimental results onCadik and CSDD database show that the proposed algorithm is superior to some existing typical grayscale algorithms, which can effectively preserve the contrast and structural information of the original image and improve the computational efficiency. The visual perception is natural for the grayscale image, whose subjectivity and objective evaluation results are optimal.