MAEA-DeepLab: A semantic segmentation network with multi-feature attention effective aggregation module

To realize the low cost of network training, the computational complexity is greatly reduced while maintaining high precision. A semantic segmentation network with multi-feature attention effective aggregation module(MAEA) is proposed: MAEA-DeepLab. A 16 stride low-resolution feature map for down-sampling is adopted in the encoder’s network backbone, and high-level features are obtained. The decoder makes full use of the feature's spatial attention mechanism through the MAEA module, effectively aggregates multiple features, and obtains high-resolution features with strong semantic representation. Then the ability of the decoder to recover important details is effectively improved, and high-precision segmentation is achieved. Multiply-adds in MAEA-DeepLab is 943.02B, only 30.9% of the DeepLabV3+ architecture, which greatly reduces the computational complexity. The architecture is not pre-training on the COCO dataset. It performs semantic semantic segmentation Benchmark tests on the test set of with PASCAL VOC 2012 dataset and CityScapes dataset with only two RTX 2080ti GPUs, and the mlOU scores reach 87.5% and 79.9%, respectively. The experimental results show that good semantic segmentation accuracy is achieved with low computational cost in MAEA-DeepLab.