Many modern applications—such as autonomous vehicles, medical imaging, and security surveillance—rely heavily on accurate and efficient image detection systems. Although recent advances in computer vision and deep learning have led to significant improvements, several challenges still persist, including limited detection accuracy, high computational overhead, and difficulties in achieving real-time performance. To address these challenges, this work focuses on leveraging lightweight neural network architectures combined with well-optimized machine learning techniques. The proposed approach is primarily built on transfer learning using pre-trained models, complemented by enhancements such as non-maximum suppression and anchor-based object detection to improve localization and classification performance. In addition, model optimization strategies, including quantization and pruning, are employed to reduce memory usage and computational complexity without compromising accuracy. The effectiveness of the proposed system is validated using widely recognized benchmark datasets such as COCO and PASCAL VOC, demonstrating its reliability and scalability across different deployment scenarios. Furthermore, the results show notable improvements in inference speed, recall, and detection precision, making the system particularly suitable for resource-constrained environments. Overall, the findings provide a strong foundation for developing robust, efficient, and scalable image detection solutions applicable across a wide range of real-world industries.