Abstract—: “MODI lipi” is one of the Indian ancient scripts and is as yet unrecognized script. It is not in use today, but it has importance for historical researchers of not only the ancient Maratha history but history in different regions of India. The recognition of MODI demands transform invariant approach, as MODI documents are deformed severely. Invariant handwritten character recognition is accomplished in the past by employing feature extraction methods, yet there is a scope to improve the results under global transformations. At present, convolutional neural network exhibits only local transform invariance impulsively by convolution-pooling architecture and data augmentation. To achieve global invariance for MODI recognition, the proposed classification framework used CNN-based transfer learning and a global feature extractor histogram of oriented gradient. Additionally, the criterion based on principal component analysis and confusion matrix are introduced to choose the invariant feature and to find classes responsible for poor recognition rate. The proposed classifiers are trained on a self-created handwritten MODI character dataset and tested on transformed MODI dataset. The results showed that the proposed framework is effective to recognize MODI handwritten characters under transformations without data augmentation and network alteration. © 2022, Pleiades Publishing, Ltd.