CSADnet: Attention-Based Semantic Segmentation Method for Spraying Scene Images of Building Gables

Accurate and efficient detection of spray-painted regions and boundary delineations on building gables is crucial for promoting automated construction processes. External wall spraying robots rely on robust environmental perception. However, existing semantic segmentation techniques frequently suffer from low detection accuracy and inadequate feature representation under real-time constraints. Traditional convolution-based deep methods also exhibit limitations in handling complex boundary delineations. To address these challenges, we propose an innovative framework called Channel-Spatial Attention Deeplabv3+ Based Network (CSADnet). CSADnet enhances context modeling and global feature representation through a channel-spatial attention mechanism, effectively overcoming the commonly observed low recognition rate in current methods. A cross-spatial feature fusion module further integrates local and global information, thus improving segmentation performance. To tackle the issue of imbalanced training data, we introduce a weighted cross-entropy loss function that stabilizes model convergence. Additionally, we adopt a Chebyshev distance–based boundary error evaluation metric for a more intuitive assessment of spraying boundary accuracy. Experimental results on a real-world building image data set demonstrate that CSADnet outperforms existing state-of-the-art approaches. Compared with the benchmark network, CSADnet achieves a 2.44% improvement in mean intersection over union and a 1.97% increase in mean pixel accuracy, and average and maximum boundary errors are reduced by 9.15% and 9.80%, respectively. These findings highlight the remarkable performance and innovative merits of CSADnet in construction robotics, paving the way for more efficient and high-precision gable painting operations.