Abstract:

Remote sensing image processing technology based on deep learning can prospectively be used to determine the characteristics of large numbers of remote sensing image data and complex scenes. However, deep-learning algorithms in remote sensing image processing have certain shortcomings, e.g., the popular DeepLabv3+ network has slow fitting speeds, inaccurate edge target segmentation, inconsistencies, and holes in large-scale target segmentation. We therefore proposed a method for introducing a Dual Attention Mechanism Module (DAMM) to DeepLabv3+ to address the above deficiencies. We designed two different network models that connected the DAMM structure to the Atous Spatial Pyramid Pooling (ASPP) layer in series or parallel. In the serial connection method, the feature map was first sent to the DAMM and then passed through the ASPP structure. Furthermore, the feature map was defused with middle-low layer feature information through the decoder layer and restored to the original image resolution. In the parallel connection method, the DAMM and ASPP layers processed the feature map extracted from the backbone network in parallel and subsequently fused the processed feature map information. The mixed feature map was restored to its original resolution by the decoder. The two improved methods were verified by the INRIA Aerial Image high-resolution remote sensing dataset. The results showed that both the series and parallel methods could effectively improve the shortcomings of Deeplabv3+. The experimental results showed that the parallel network had superior performance, and improvements in the original network defects were more obvious. The parallel method achieved a higher score [85.44% Mean Intersection Over Union (MIOU)] in the test dataset, which was 1.8% higher than Deeplabv3+. And the serial network increased by 1.12% compared to Deeplabv3+. The effects of the position and channel attention mechanisms in the DAMM structure were also determined. The ablation study results showed that the channel and position attention mechanisms improved the performance of the Deeplabv3+ model. In the test set, the channel and position attention mechanism mIoU increased by 0.95 and 1.32%, respectively. The experiments revealed that the position attention mechanism had a greater effect on edge target segmentation, the channel attention mechanism had a greater effect on large-scale hole phenomena, and the channel and position attention mechanism promoted network fitting speed in training. The proposed improved DeepLabv3+ algorithm can provide a scientific basis and reference for semantic segmentation of big data remote sensing images.

Key words: remote sensing image, deep learning, DeepLabv3+, attention mechanism, semantic segmentation