Image semantic segmentation approach based on DeepLabV3 plus network with an attention mechanism

Abstract

Image semantic segmentation is a technique that distinguishes different kinds of things in an image by assigning a label to each point in a target category based on its "semantics". The Deeplabv3+ image semantic segmentation method currently in use has high computational complexity and large memory consumption, making it difficult to deploy on embedded platforms with limited computational power. When extracting image feature information, Deeplabv3+ struggles to fully utilize multiscale information. This can result in a loss of detailed information and damage to segmentation accuracy. An improved image semantic segmentation method based on the DeepLabv3+ network is proposed, with the lightweight MobileNetv2 serving as the model's backbone. The ECAnet channel attention mechanism is applied to low-level features, reducing computational complexity and improving target boundary clarity. The polarized self-attention mechanism is introduced after the ASPP module to improve the spatial feature representation of the feature map. Validated on the VOC2012 dataset, the experimental results indicate that the improved model achieved an MloU of 69.29% and a mAP of 80.41%, which can predict finer semantic segmentation results and effectively optimize the model complexity and segmentation accuracy.