A comprehensive study on optimization techniques for AMR robots recognition models

Abstract

In the process of computing on digital computers and related devices, problems arise in the representation, operation, and storage of numerical values. Quantization and pruning are common methods to address these issues. Using quantization, real-valued numbers can be converted to discrete numbers, thus reducing computational costs. In recent years, neural network models(such as image classification, object detection, semantic segmentation, natural language processing, etc.)have developed rapidly in many fields. However, in specific domains such as computer vision and natural language processing, neural network models often have high computational requirements, making the quantization process essential. Quantization technology can compress parameters by converting floating point parameters in the model into low-precision integer parameters. This reduces the storage space required for the model and the time it takes to load the model. Autonomous Mobile Robots (AMR) represent a practical application of recognition models, where model compression becomes essential during deployment due to constraints imposed by hardware limitations. This report investigates the core techniques utilized in optimizing neural network models and transformer models for the deployment of AMR systems, with a specific emphasis on quantization and pruning methodologies, and widely adopted optimization frameworks.