An attempt was made to create an Improved Masked Face (IMF) dataset covering the various aspects of pose, occlusion, expressions, illumination, and multiple real masks that vary in texture and color along with nose coverage. The existing masked face datasets were created using simulated masked face images, and the trained models have challenges under real situations. The proposed dataset has images of the same person wearing different types of masks under varied illumination which added the diversity of the dataset. The developed IMF dataset is unique covering the different possibilities in which a mask can be worn. The proposed dataset has a total of 54,408 face samples belonging to 13 persons. We also propose a light weight improved IMF ConvNet as a deep learning classification model developed on CNN architecture and the MTCNN face detector which reported an accuracy of 99.9% on the IMF dataset.