Received Signal Strength Indicator (RSSI) fingerprinting has emerged as a promising solution for indoor localization, offering a practical approach to position estimation in GPS-denied environments. However, environmental dynamics, intrinsic signal noise, and limited computational resources present significant challenges to traditional methods. In this study, we propose a lightweight and optimized model, MobileNetV2-mini++, for Wi-Fi RSSI-based indoor localization. The proposed architecture leverages separable convolutions, adaptive learning rate scheduling, and overfitting mitigation strategies to strike an effective balance between accuracy, speed, and resource consumption. Hyperparameters were carefully optimized through grid-based tuning, and a controlled random-noise augmentation method (±3 dBm) was applied to improve robustness against signal fluctuations. For fair benchmarking, AlexNet and ResNet were selected as representative classical and modern CNN architectures. A real-world dataset comprising over 110,000 RSSI samples collected from 35 reference points within the Faculty of Geography at the University of Tehran was used for model evaluation. On augmented data, the model achieved an accuracy of 88.39%, a precision of 90.29%, and an F1-score of 88.08%. Furthermore, in noisy real-world conditions, MobileNetV2-mini++ demonstrated superior robustness compared to baseline architectures, achieving the highest accuracy of 62.77%. The model also reduced the localization error to 0.7121 units. These results indicate that MobileNetV2-mini++, while maintaining architectural simplicity, exhibits strong resilience to environmental challenges and can serve as an effective solution for real-time indoor positioning systems. Future directions include multimodal data integration, intelligent noise handling, and deployment on mobile devices.