Autocorrelation-Aware Aggregation Network for Salient Object Detection of Strip Steel Surface Defects

In recent years, salient object detection (SOD) has made great progress in natural scene images (NSIs), but SOD of strip steel defect images (SDIs) in industrial scenes is still an open and challenging problem. Existing detection methods are difficult to segment different types of defects with clutter and shallow contrast. Therefore, we propose a novel autocorrelation-aware aggregation network (A3Net) for SOD of strip steel surface defects. First, we use a general residual attention mechanism to enhance the encoder features and accelerate the convergence of the model. In the decoder stage, we propose a global autocorrelation module (GAM) to explore semantic information cues of high-level features to locate and guide low-level information. Then, we deploy the scale interaction module (SIM) to realize the fusion and interaction of feature information between different layers. Finally, we design a local autocorrelation module (LAM) to further refine the edge details of salient objects. We conduct detailed and rich experiments on the public strip steel surface defects dataset, which proves that our method is consistently superior to the state-of-the-art methods. In addition, we build a new challenging strip SDI dataset with multiple defect types for the SOD task, which contains 4800 images with pixel-level annotations. Our dataset and code are available at https://github.com/VDT-2048/A3Net.