关键词: 图像处理, 语义分割, DeepLabV3+, 通道剪枝, 拆分三重注意力, 细粒度上采样

Abstract: The current high-precision semantic segmentation model requires huge computing resources, so it is difficult to deploy on embedded platforms with limited hardware storage and computing power. Aiming at this issue, an image semantic segmentation model based on feature fusion and attention mechanism is proposed. Firstly, the model based on DeepLabV3+ is optimized and the MobileNetV2 backbone network is lightened using channel pruning. Secondly, the Splittable Triplet Attention (STA) is introduced to the lightweight model to improve the internal dimensional correlation of the feature map. Finally, fine-grained up-sampling modules are added in the decoding part to improve the edge detail information. In the experiments on Pascal VOC 2012 and cityscapes datasets, the parameter number of the proposed algorithm is only 4.15×106, the number of floating-point operations is 10.23 GFLOPs, and the average intersection ratio is 70.98% and 72.26% respectively. The results show that the model achieves a good balance among computing resources, memory consumption and accuracy.

Key words: image processing, semantic segmentation, DeepLabV3+, channel pruning, splittable triplet attention, fine-grained upsampling