MobileNet 기반 DeepLabv3+의 경량화 방법

초록

Recent advances in autonomous driving, medical image analysis, and surveillance technologies have led to an increasing demand forreal-time image segmentation. Consequently, research on real-time segmentation methods has been actively pursued, with deeplearning-based approaches demonstrating high performance metrics. However, the complex architectures, massive computational cost,and large parameter counts inherent in artificial neural networks impose significant constraints when deploying them on resource-limitedembedded platforms. To address this issue, various attempts have been made, such as employing lightweight classification networks likeMobileNet as backbone networks for segmentation models. Nevertheless, segmentation networks still involve substantial computationaloverhead and parameter size, making them difficult to apply in resource-constrained environments. Therefore, even when lightweightclassification networks are used, further compression is necessary. In this paper, we propose a lightweight approach for DeepLabv3+which MobileNetV3-Large without the last expansion layer as the backbone network. To demonstrate the effectiveness of the proposedmethod, we compare it with several DeepLabv3+ architectures whose backbone networks are MobileNetV3-Large, MobileNetV3-Small,MobileNetV3-Small without the last expansion layer, or MobileNetV3-Large compressed by hyperparameter tuning. Compared to the bestperforming architecture with MobileNetV3-Large as the backbone network, the structure with the proposed method can reduce theparameters by about 52% with only a 2.7%p performance decrease. Therefore, the proposed method effectively achieves the trade-offbetween performance and neural network size, which confirms that it can be practically utilized in resource-constrained environments.

키워드