Model optimisation and energy efficiency in deep learning based image classification on edge devices

This paper makes a significant contribution to the body of knowledge in the field by analyzing in detail the various applications of deep learning techniques on edge devices with a focus on performance and energy efficiency. The remarkable advances in deep learning algorithms in recent years have led to both high computational power requirements and high energy consumption of the resulting models. This has necessitated further research and development on energy efficiency. The focus is primarily on the integration of cutting-edge lightweight and contemporary architectures, notably MobileNet, EfficientNet, and ShuffleNet. These architectures are extensively utilized across a broad spectrum of applications, extending from mobile devices to innovative smart sensors that interact with various environments. This crucial work, characterized by its significance, has been meticulously carried out through systematic and comprehensive testing. The hardware employed in this process is notably diverse, encompassing platforms such as the AMD NPU, Google Coral TPU, and NVIDIA Jetson Nano, each contributing to the validation and effectiveness of the architectures. Each test was carefully planned and executed according to specific criteria, and the results were meticulously analyzed. The best performance results are supported by reliable data and contribute to setting new standards in the field of deep learning. The findings of this study, combined with this thesis, establish a vital groundwork for deep learning technologies to become increasingly accessible, particularly on edge devices, while also significantly enhancing their efficiency. Recent developments within the field of deep learning have an extensive array of applications, extending from mobile devices to intelligent sensors, which possess the tremendous potential to influence and shape the technologies of the future. Therefore, this thesis not only conducts a comprehensive analysis of existing applications in great detail but also offers essential contributions to future research and development processes, paving the way for innovations that could transform numerous industries and improve human-computer interactions in unprecedented ways. The study serves as a vital resource that supports the ongoing development and enhancement of deep learning techniques and their various applications across multiple fields. The significant findings derived from this study can certainly be advanced further through collaborative efforts that unite engineers and academics strategically. Their teamwork will greatly increase the overall effectiveness of AI applications, particularly on edge devices, which are increasingly important in today's technology ecosystem. Ultimately, this comprehensive research not only compiles the existing body of knowledge but also charts a well-defined roadmap for the future trajectory of deep learning. Additionally, it lays a sturdy foundation that is crucial for the further expansion of this dynamic field. This progressive approach paves the way for subsequent studies and contributes significantly to refining capabilities that are essential for meeting the evolving needs of the industry as a whole. The insights provided by the study help foster innovative applications of deep learning, thereby leaving a lasting impact on the evolution of the technology. As a result, this research goes beyond compiling existing knowledge and provides a roadmap for the future of deep learning, as well as a solid foundation for the expansion of the system's ecosystem. In this comprehensive and thorough study, a series of carefully designed and detailed experiments were meticulously conducted, focusing on advanced model compression techniques such as pruning and quantization. Furthermore, the comprehensive and in-depth integration of a broad array of innovative lightweight architectures, which prominently include MobileNet, EfficientNet, and ShuffleNet, was explored in great detail. These targeted and specialized methodologies were specifically designed to guarantee not only the efficient operation of deep learning models but also to ensure high performance when they are deployed across a diverse range of devices. This, in turn, significantly enhances their usability and effectiveness across multiple platforms, making them suitable for various applications and use cases. The exploration of these architectures highlights their importance in the evolving landscape of technology. Through systematic and rigorous testing across a diverse yet representative array of hardware platforms, including but not limited to AMD NPU, Google Coral TPU, and NVIDIA Jetson Nano, the researchers were able to ascertain the best performance results achievable with these methods. This thorough process led to the collection of reliable and impactful data that underscores the effectiveness of the findings. By undertaking this extensive and in-depth research effort, the thesis makes a significant contribution to enhancing both the accessibility and efficiency of cutting-edge deep learning technologies. This advancement, in turn, facilitates their application on edge devices in a more effective and scalable manner, ultimately promoting a wider adoption of such technologies across various sectors. Innovations in deep learning possess impressive and transformative potential to profoundly influence the future landscape of numerous advanced technologies, spanning from cutting-edge mobile devices designed for seamless user experiences to complex and sophisticated smart sensors that enhance various applications in our daily lives. Given this significant impact, this thesis not only delivers a comprehensive analysis of existing applications and their various implications but also illuminates the promising avenues for future research and continuous development that lie ahead in this dynamic and exciting field. By exploring these avenues, we can better understand the transformative power of deep learning and the opportunities it presents for innovation across multiple domains. This work is a critical resource supporting the development of deep learning techniques and applications. These findings can be further developed with the collaboration of engineers and academics to increase the effectiveness of AI applications on edge devices. In conclusion, this research goes beyond compiling existing knowledge and provides a roadmap for the future of deep learning.