A Hybrid Stacked CNN Model with Weighted Average Ensembling for Effective Face Recognition

Abstract

The discipline of computer vision has given a lot of attention to facial recognition. Automated face recognition is extensively employed in various practical scenarios, including systems for streamlining immigration checkpoints, intelligent monitoring of visual data, and authentication of personal identity. Depending on the situation, it may be divided into the two separate duties of facial verification and face identification. This study proposed a hybrid stacked CNN model for face recognition system. The models have mastered the art of making their own inferences. The models are then combined to predict a class value using a cutting-edge method called weighted average ensembling. A more accurate estimate should be produced by the new assembly process. Pre-trained CNN models are used in our proposed method: AlexNet, Resnet50V2, and VGG-19, Yolov5, VGG-16 and ResNet50. When applied to Tufts dataset images, our suggested model successfully achieved 98.05% accuracy. We have also used the Discrete Wavelet Transform (DWT) method for denoising, SegNet for Image segmentation for better performance of the model proposed.