Face detection is a prominent research domain in the field of digital image processing. Out of various algorithms developed so far, Viola–Jones face detection has been highly successful. However, because of its complex nature, there is need to do more exploration in its various phases including training as well as actual face detection to find the scope of further improvement in terms of efficiency as well as accuracy under various constraints so as to detect and process the faces in real time. Its training phase for the screening of large amount of Haar features and generation of cascade classifiers is quite tedious and computationally intensive task. Any modification for improvement in its features or cascade classifiers requires re-training of all the features through example images, which are very large in number. Therefore, there is need to enhance the computational efficiency of training process of Viola–Jones face detection algorithm so that further enhancement in this framework is made easy. There are three main contributions in this research work. Firstly, we have achieved a considerable speedup by parallelizing the training as well as detection of rectangular Haar features based upon Viola–Jones framework on GPU. Secondly, the analysis of features selected through AdaBoost has been done, which can give intuitiveness in developing more innovative and efficient techniques for selecting competitive classifiers for the task of face detection, which can further be generalized for any type of object detection. Thirdly, implementation of parallelization techniques of modified version of Viola–Jones face detection algorithm in combination with skin color filtering to reduce the search space has been done. We have been able to achieve considerable reduction in the search space and time cost by using the skin color filtering in conjunction with the Viola–Jones algorithm. Time cost reduction of the order of 54.31% at the image resolution of 640*480 of GPU time versus CPU time has been achieved by the proposed parallelized algorithm.