This paper presents a real-time thermal face detection and recognition system based on an enhanced Multi-task Cascaded Convolutional Neural Network (MTCNN) framework. Unlike visible-light methods, thermal imaging introduces domain-specific challenges such as low spatial resolution, high noise, and intensity variance due to temperature fluctuations. To address these, we propose a dedicated preprocessing pipeline including normalization, contrast enhancement, and channel replication to adapt single-channel thermal images for CNN-based processing. The modified MTCNN is fine-tuned on thermal datasets to accurately detect facial regions and landmarks. Aligned faces are then processed through a thermal-optimized feature embedding network trained with triplet loss to produce identity-preserving descriptors. Recognition is performed using a lightweight classifier over the feature space. The system is optimized for real-time performance using GPU acceleration and quantized inference. Experimental results on publicly available thermal face datasets demonstrate the effectiveness of our approach in terms of detection accuracy, recognition rate, and processing speed, making it suitable for surveillance and biometric applications under low-light or no-light conditions.