Recent advancements in artificial intelligence and computer vision have positioned deep neural networks (DNNs) as powerful tools for improving the performance of complex systems, such as image and video recognition. These technologies enable the implementation of sophisticated systems like facial recognition, autonomous vehicle navigation, and natural language processing. However, the inference process in DNNs requires significant computational resources, presenting challenges when deployed on edge devices with limited resources. This thesis introduces and eva‎luates a collaborative inference method based on CNNs (Convolutional Neural Networks) across multiple edge devices, aiming to enhance inference speed while maintaining model accuracy, all within the constraints of the limited hardware resources of edge devices. CNNs are a type of DNNs that are usually used in image processing applications. The proposed method involves partitioning of the input data of neural network model and distributing the computational tasks across several edge devices. This partitioning is designed to reduce resource consumption and improve processing speed. A key innovation in this research is the parallel implementation of pooling layers across devices, rather than performing these operations centrally. This adjustment in the processing structure significantly reduces the need to aggregate outputs at a central device, thereby markedly increasing inference speed. Experimental results demonstrate that the proposed method achieves substantial improvements in inference speed and overall efficiency compared to traditional and similar approaches. This method, compared to similar approaches, has increased inference speed by an average of 30% while maintaining model accuracy. These advancements have significant implications for a wide range of applications, including surveillance systems, the Internet of Things (IoT), and intelligent systems such as autonomous vehicles.

مسعودرضا هاشمي

امير خورسندي كوهانستاني , سمانه حسيني