Lightweight Semantic Segmentation Network Based on DeepLabV3+

Abstract

Embedded mobile devices have limited computing power and insufficient running memory, and it is difficult to deploy high-precision, high-complexity and time-consuming semantic segmentation models. We propose a lightweight semantic segmentation model based on DeepLabV3+. This model optimizes the original DeepLabV3+ model from the perspective of reducing the amount of parameters and ensuring segmentation accuracy. The original backbone network is replaced by the MobileNetV2 network with lower parameters and computational complexity to speed up model inference. We design a 3-branch parallel structure and introduce a Semantic Embedding Module (SEB) to enhance low-level feature map semantic information and pixel point feature representation. The model adds a recurrent cross-attention mechanism module (RCCA) to capture the global correlation of all pixels and obtain dense contextual information. The model achieves 74.81% Mean IoU on the mixed dataset consisting of PASCAL VOC 2012 and Semantic Boundaries Dataset, with a parameter size of 8.27MB. The comprehensive performance of the model is better than that of networks such as SegNet, BiSeNetV2 and ENet, and a good balance is achieved between segmentation accuracy and model complexity.