Face recognition is a vital aspect of computer vision and biometric technology, with applications ranging from security. It is the most critical research direction for identifying criminal activities. The problem of face detection under arbitrary occlusion has become a major concern for social security due to the use of surveillance systems to detect crimes. In complex environments, many researchers use ML-based techniques for face recognition, but there has been no satisfactory recognition accuracy for recognizing faces. In this paper, a novel deep learning-based SH-GAN is proposed for efficient regeneration and recognition of human faces. Initially, the masked face images are gathered from publicly available dataset and these images are pre-processed using bilateral filter to remove the noisy artifacts. Then, the noise-free images are fed in the DL-based U-net for segmenting the masked region to create overlaid images. The segmented mask and overlaid image are given as input to the SIFT integrated HOG based GAN for regenerating the facial images based on the ground truth. Additionally, in SH-GAN the regenerated images are identified as authorized and unauthorized (unknown) faces. The experimental results of the proposed model are assessed using specific metrics like accuracy, F1 score, dice index and jaccard index. From this analysis, the proposed SH-GAN attains the overall accuracy of 98.14% in the recognition of facial images. The proposed SH-GAN framework increases the overall accuracy of 3.99%, 7.94% and 27.17% for Face mesh model, MFNet and Haar Cascade technique respectively.