In the field of medical research on skin diseases, machine learning categorization approaches are becoming increasingly prominent. The classification process frequently makes use of the extraction of visual features based on statistical analysis. Statistically-based image feature extraction approaches are not as effective as deep models, which are learned skin images. Examples of deep models include convolutional neural networks, Mobilenet, Resnet, Densenet, and GoogleNet. Deep models are employed to locate features more efficiently. There are several concerns with deep features, including those that pertain to the curse of dimensionality and data sparsity issues. The spectral notion that is utilized for the purpose of locating the lower manifold subspace of deep features is successfully handled by subspace learning techniques, which are the most successful methods for dealing with the curse of dimensionality. There was also a reduction in the data sparsity impact that occurred throughout the process of translating a high-dimensional subspace into a lower manifold space. By locating the principal components, also known as principal Eigenvectors, it is possible to acquire equivalent spectral features. In comparison to the dimensions of deep features, the number of primary components is substantially lower. The process of mapping the deep features onto the spectral (or Eigen) lower manifold space is referred to as the spectral space. This process is effective in preserving the information contained in the skin image data. In order to achieve robust skin disease classification, the work that is being proposed expands the deep-based classifier techniques by using the spectrum notion. Through the utilization of the benchmarked dermoscopic and clinical image datasets, the experimental investigation demonstrates the effectiveness of the offered methodologies.