Multi-Object Extraction of Remote Sensing Images Based on Improved DeepLabv3+

In the context of multi-object extraction from high-resolution remote sensing images., issues such as class imbalance leading to accuracy degradation and insufficient utilization of feature information persist. This paper proposes an enhanced DeepLabv3+ network to achieve multi-object recognition and extraction for high-resolution remote sensing imagery. The modified DeepLabv3+ replaces Xception with ResNet50 as the backbone network to enrich semantic information. Additionally., a coordinate attention mechanism and a triple attention mechanism are incorporated to fully exploit feature information and strengthen object extraction capabilities. Moreover., the Lovász-Softmax loss function is employed to mitigate accuracy reduction caused by imbalanced sample distribution. The proposed model is tested on the Potsdam dataset to validate its feasibility. Furthermore., the model is compared with OCNet., DenseASPP., DeepLabv3+., AFF-UNet.,BSNet.,LSP-Net., and others on both the Potsdam and GID-15 datasets to evaluate its effectiveness. Experimental results demonstrate that the proposed model achieves well-defined target structures with smooth edges and strong feasibility., outperforming the listed comparison networks in terms of MIOU.,Recall., and OA on both datasets.