The development of a generative model brings significant changes in the research, including object detection, text detection, and recognition in scene images. At the same time, they pose unexpected challenges to achieving stable and reliable performance of the method. This study presents work for the classification of real and deepfake images such that an appropriate model can be chosen for improving the detection performance. We present a new model for classifying deepfake and real images based on the fact that the generative models introduce a kind of impurity, which results in loss of edges, artifacts, deformation, and tampering with the content of the images compared to the original images, irrespective of the generative models. However, the impurity in the output images cannot be predicted as generative models change, which makes classification more challenging and interesting. To alleviate this challenge, the proposed work integrates the strengths of cross domains. We propose the Non-Correlative Feature Interaction Matrix (NCFIM), a multi-modal framework for image-based deepfake image classification. NCFIM integrates features from raw pixels, DCT, FFT, and entropy statistics, capturing inconsistencies across domains for robust and interpretable detection. The CUDA-optimized pipeline with Automatic Mixed Precision (AMP) ensures efficient deployment. To demonstrate the efficacy of the proposed classification model, experiments are conducted on collected deepfake images by mixed generative models and deepfake detection, text detection, and recognition before and after classification and detection. Experiments on the CIFAKE and ICDAR-GAN datasets demonstrate the efficacy of the proposed method in both classification accuracy and discriminative reliability, achieving a new state-of-the-art ROC-AUC of 0.9986 on CIFAKE and 0.7780 on ICDAR-GAN. Additionally, an average improvement of 0.0783 in F1 score for text detection was observed after classification across various methods on the ICDAR-GAN dataset.