目的 主成分分析网络(PCANet)能提取图像的纹理特征,线性判别分析(LDA)提取的特征有类别区分性.本文结合这两种方法的优点,提出一种带线性判别分析的主成分分析网络(PCANet-LDA),用于视网膜光学相干断层扫描(OCT)图像中的老年性黄斑变性(AMD)、糖尿病性黄斑水肿(DME)及正常(NOR)这3类的全自动分类.方法 PCANet-LDA算法是在PCANet的基础上添加了LDA监督层,该层加入了类标签对特征进行监督投影.首先,对OCT视网膜图像进行去噪、二值化及对齐裁剪等一系列预处理,获得感兴趣的视网膜区域;然后,将预处理图像送入一个两层的PCA卷积层,训练PCA滤波器组并提取图像的PCA特征;接着,将PCA特征送入一个非线性输出层,通过二值散列和块直方图等处理,得到图像的特征;之后,将带有类标签的图像特征送入一个LDA监督层,学习LDA矩阵并用其对图像特征进行投影,使特征具有类别区分性;最后,将投影的特征送入线性支持向量机(SVM)中对分类器进行训练和分类.结果 实验分别在医院临床数据集和杜克数据集上进行,先对OCT图像预处理进行前后对比实验,然后对PCANet特征提取的有效性进行分析,最后对PCANet算法、ScSPM算法以及提出的PCANet-LDA3种分类算法的分类效果进行对比实验.在临床数据集上,PCANet-LDA算法的总体分类正确率为97.20％,高出PCANet算法3.77％,且略优于ScSPM算法;在杜克数据集上,PCANet-LDA算法的总体分类正确率为99.52％,高出PCANet算法1.64％,略优于ScSPM算法.结论 PCANet-LDA算法的分类正确率明显高于PCA-Net,且优于目前用于2D视网膜OCT图像分类的先进的ScSPM算法.因此,提出的PCANet-LDA算法在视网膜OCT图像的分类上是有效且先进的,可作为视网膜OCT图像分类的基准算法.

Image understanding is based on image semantic segmentation. By employing patch shape complexity index and Euler number, an image segmentation algorithm ( GM ) based on over-segmentation, patch merging and graph model is presented. Following a common technical tactic that an image is segmented after smoothing, 16 algorithms for image segmentation are presented by combining each of the 4 filter methods including Bilateral Filtering ( BiF ) , Edge Preserving Smoothing (EPS), structure extraction via Relative Total Variation (RTV) and Rolling Guidance Filter ( RGF ) , with each of the 4 segmentation algorithms including Threshold Segmentation ( ThrS) , Active Contours without Edges ( ACnE) , Particle Swarm Optimization ( PSO) , and GM. To evaluate the efficiency of the algorithms, the comparative experiments are conducted, and two evaluation indexes, Corrected Rand Index ( CRI ) and Jaccard Index ( JI ) , are used in the experiments. The experimental results show that the algorithm which segments an image with GM after RTV has the best performance from CRI and JI. T-test also shows that it is statistically significant that the segmentation effect of GM algorithm is improved by using image smoothing.

Extraction of multiple resolution patches from image is helpful to image semantic segmentation and scene understanding. Based on the analysis of the algorithm for structure extraction which realized via relative total variation and its parameter value, a semi-automatic algorithm for the extraction of multiple resolution patches is presented in this paper. We put forward a piecewise function of iteration step length and give an optimal stop criterion for the iterative algorithm. Our experiment results show that our method is of general applicability and feasibility. Key-Words: Multiple resolution patches, Extraction of patches, Image segmentation, Optimal partition, Annotation, Semantic segmentation, Smoothing.