Semantic Segmentation Model for Edge Devices Based on Model Compression and Model Acceleration

Abstract

This study aims to adopt different models and apply various model compression and acceleration techniques to improve the performance of segmentation and computing efficiency of the semantic segmentation models for small objects such as safety guardrails. Due to the scarcity of data, caused by restrictions of construction site safety regulations and difficulty of data labeling, this study adopts the method of data augmentation to assist the training process of the model. In addition, in response to the model with hardware performance and a large number of model parameters, it is found that using input images of different sizes for different models can ensure its segmentation performance and successfully perform guardrail identification according to the experiments. As a result, all models achieve above 0.54 in IoU. In this study, Ghost Module is chosen as the acceleration method, and experiments have confirmed that this acceleration method can help improve the computing efficiency and allow the performance of segmentation of the model up to an IoU of 0.65. Although running on edge devices cannot achieve the level of real-time segmentation, after model acceleration, the time required for an image is still significantly decreased by more than 110 percent. Also, since the guardrail is a static object, there is no need for a fast identification frequency. Finally, in order to further reduce the computational complexity of the model, this study uses model pruning to compress the overall model size. According to the results of the experiments, it is found that there is indeed a problem of redundant weights in the model. After removing a certain degree of redundant weights by the L1 norm and adopting fine-tuning, it can effectively improve the model's ability to segment guardrails