Improved Lightweight DeepLabv3+ Algorithm Based on Attention Mechanism

Abstract

DeepLabv3+ has a wide range of applications in autonomous driving, geographic information systems, etc. However, its deployment on the mobile terminal faces a trade-off between model size and accuracy. Consecutive downsampling operations also result in a great loss of detail information. To solve these problems, this paper proposes an improved algorithm based on DeepLabv3+. Firstly, backbone is replaced by MobileNetv2 to reduce the size of the model; Secondly, the improved Atrous Spatial Pyramid Pooling module is proposed to augment the segmentation result while reducing the parameters. The performance is further ameliorated by applying attention mechanism; Finally, through refining decoder module, the proposed network makes up for lost detail information. Experiment shows that the algorithm achieves an mIoU of 73.31% on the validation set of the PASCAL VOC2012 dataset. Compared with typical algorithms, proposed algorithm has a better effect on trade-off between model size and accuracy.