YOLO V8 Network Lightweight Method Based on Pruning and Quantization

Abstract

With the development of deep learning, convolutional neural networks have been widely used as core algorithms in fields such as computer vision, natural language processing, and speech processing, surpassing the performance of traditional algorithms. However, its complex structure and massive computational load make many algorithms dependent on GPU implementation, which limits its application on mobile devices with limited resources and high real-time requirements. This article explores in detail the lightweighting process of the YOLO v8 model, aiming to optimize its performance in resource constrained environments by combining BN layer scale factor pruning and int8 quantization techniques under the TensorRT framework. Firstly, using the scale factor in the BN layer as an evaluation criterion, redundant channels in the network are identified, effectively reducing the number of model parameters and computational requirements; Then deploy the model in NVIDIA's TensorRT framework for int8 quantization. This process compresses the floating-point operations of the model into low bit int8 operations, significantly reducing computational complexity and memory usage; And validated through COCO dataset experiments during the process. The results show that YOLO v8, after pruning and quantization, can achieve a compression ratio of 64.2% with a loss detection accuracy of about 4%. This study provides a comprehensive solution for the efficient deployment of deep learning models, especially in application scenarios that require balancing performance and resource efficiency, which has important practical significance. Future work will focus on further optimizing pruning and quantization strategies to adapt to a wider range of network architectures and application requirements.