An accurate and resource-efficient network for surface anomaly detection via enhanced downsampling and activation representation

In deep learning-based anomaly detection, performance degradation often results from feature loss caused by downsampling and the limited nonlinear representation capacity of conventional activation functions. Moreover, improving detection accuracy often demands substantial computational resources, thereby hindering practical deployment in real-time and resource-constrained collaborative multi-robot inspection settings. To address these challenges, this paper proposes DEANet, a memory-efficient and real-time anomaly detection methodology with four key components, designed to achieve high accuracy at low computational cost. First, a lightweight feature aggregation neck improves feature fusion efficiency while reducing computational overhead. Second, a contextual feature extraction module leverages environmental semantics to enhance both detection and localization accuracy. Third, to alleviate the feature degradation introduced by hierarchical downsampling, two enhancement modules are designed to facilitate a better trade-off between accuracy and computational cost. Fourth, we propose a parameterized activation function (ACLU) that enhances the network's nonlinear representational capacity. ACLU achieves higher accuracy and demonstrates faster convergence compared to recent advanced activation functions. Experiments on three benchmark datasets confirm that DEANet achieves state-of-the-art accuracy with only 2.8 million parameters, while reducing training data demands by 70%, parameter count by 87.8%, and computational cost by 92.6%, demonstrating its strong efficiency–performance trade-off under resource-constrained conditions. Edge-computing deployment tests validate DEANet's real-time performance at 52.1 FPS. These results highlight DEANet's practicality and scalability for deployment in real-world, resource-constrained settings. The source code will be available at https://github.com/chriszxk/surface-detection.git.