An Empirical Study of Model Compression Techniques for DNN Deployment on Resource-Constrained Devices

The rapid progress in deep learning has resulted in the extensive utilization of deep neural networks (DNNs) for various computer vision applications. However, existing DNN models are computationally expensive and require a lot of memory, making it difficult to deploy them in resource-constrained devices or applications with strict latency requirements. To effectively tackle the aforementioned challenges and facilitate model deployment on edge devices, focus on model compression has emerged as a crucial area of research. This study provides a thorough review of the latest deep neural network model compression strategies to support researchers in decreasing model complexity and training expenses, increasing model deployment, speeding up model inference, and decreasing device storage space. Modern techniques like model pruning, parameter quantization, among others are covered in this work. Additionally, the study discusses the functionalities of the frameworks needed for model compression and addresses current research issues while suggesting forthcoming directions for future work.