In unsupervised classification, the subspace clustering is gaining the scope for the categorization of the more comprehensive and random image pool. In this paper, the visual and appearance features of images are evaluated independently and jointly for optimizing the subspace clustering. The normalized-image is divided into smaller blocks and extracted the visual and textural features. Entropy, Homogeneity, structural, and Edge content Features are evaluated for each block. The fuzzy rules are applied to the individual features for conducting the distinct block-adaptive hierarchical clustering. In the second level, the feature subspace is generated for exclusive features and applied to the hierarchical subspace clustering over it. After getting the cluster-segments for each image-feature and feature-subspace, the second-level fuzzy–rules are applied to assign the weights to each block. In the final stage, the image pool is processed based on this weighted poling and distance for identifying the image category. This collaborative evaluation based map performed the active clustering over the image pool. The proposed method is applied to AR, Extended-Yale, USPS, and Coil-20 Datasets. The comparative evaluation is conducted against Accuracy, NMI, and CE parameters. The proposed framework outperformed the SSC, LRR, LSR1, LSR2, SMR methods by 5.59%, 16.89%, 6.29%, 6.29%, 4.89% and 3.39% in NMI computation for AR dataset. The significant reduction in CE was achieved by 9.07%, 15.67%, 6.77%, 8.47%, 4.47% against SSE, LRR, LSR1, LSR2, and SMR methods for AR dataset. For the Extended Yale dataset, the proposed framework outperformed the existing clustering methods with 78.08% NMI and 21.11% CE. A significant higher NMI of 86.37% and least CE of 7.13% is achieved in this proposed model. For the Coil-20 dataset, the proposed model achieved 91.19% NMI and 82.83% accuracy, which is significantly better than existing methods.

在无监督分类中，子空间聚类为更全面和随机的图像池的分类提供了空间。本文对图像的视觉和外观特征进行了独立评估和联合评估，以优化子空间聚类。归一化图像被分为较小的块，并提取了视觉和纹理特征。熵，同质性，结构和边缘内容对每个块评估特征。将模糊规则应用于各个特征，以进行不同的块自适应分层聚类。在第二层中，为专用特征生成特征子空间，并将其应用于基于其的聚类子空间。在获得每个图像特征和特征子空间的聚类细分之后，第二级模糊规则用于为每个块分配权重。在最后阶段，基于此加权极化和距离来处理图像池，以识别图像类别。此基于协作评估的地图在图像池上执行了主动聚类。所提出的方法适用于AR，扩展耶鲁，USPS和Coil-20数据集。针对准确性，NMI和CE参数进行比较评估。在AR数据集的NMI计算中，所提出的框架优于SSC，LRR，LSR1，LSR2，SMR方法分别为5.59％，16.89％，6.29％，6.29％，4.89％和3.39％。相对于AR数据集的SSE，LRR，LSR1，LSR2和SMR方法，CE的显着降低实现了9.07％，15.67％，6.77％，8.47％，4.47％。对于扩展耶鲁数据集，提出的框架以78.08％的NMI和21.11％的CE优于现有的聚类方法。在此提议的模型中，NMI显着提高了86.37％，CE最低达到7.13％。对于Coil-20数据集，提出的模型实现了91.19％的NMI和82.83％的准确性，这明显优于现有方法。