This work investigates the role of canonical correlations analysis in image classification problems. Canonical correlation analysis is proposed as an alternative feature selection and reduction method for generic image classification problems. This new method is studied via various image classification problems in comparison with principal components and kernel principal components analysis. Multiple canonical correlation analysis is proposed as a new feature selection and dimension reduction algorithm for image classification problems involving multiple classes. Classification performance and relationship between the extracted image attributes and classification performance are studied by using Caltech 101 dataset.