Face recognition is considered as a promising solution for video surveillance systems. Currently, the still image-based face recognition techniques have obtained promising accuracy but detection and recognition of faces in real-time videos has becomes a challenging task. Moreover, the demand of CCTV (Closed-Circuit Television) based surveillance has increased rapidly where the quality of videos is very low. Thus, the poor-quality video, occlusion, and other conditions creates various complexities in face recognition. Currently, CNN (Convolutional Neural Network) based techniques have gained attraction from research community because these techniques have good learning capability and provide better accuracy. In this work, we have introduced CNN (Convolutional Neural Network) based scheme which uses feature extraction and feature embedding modules along with Google Net architecture to improve the learning of CNN (Convolutional Neural Network). We have incorporated histogram equalization-based image enhancement approach to improvise the quality of video frames. The proposed approach is implemented using Python 3.7. The experimental analysis shows that proposed approach achieves the accuracy as 98.55% and AUC(Area under the Curve) as 99.10% for open source datasets whereas for real-time scenarios without occlusion it achieves accuracy as 99.12%, for occlusion scenario it achieves 98.87% classification accuracy.