Multi-scale fusion convolution network with progressive dilation for real-time salient object detection of surface defects on strip steel

Accurate and efficient salient object detection (SOD) of strip-steel surface defects plays a critical role in maintaining product quality in modern industrial manufacturing. However, existing SOD methods often struggle to balance detection accuracy with inference efficiency, especially when handling complex defect patterns in real-time production environments. To address this challenge, we propose a novel framework named Multi-Scale Fusion Convolution Network with Progressive Dilation (MSFNet-PD), which is specifically designed for real-time salient defect detection. The proposed MSFNet-PD introduces a multi-scale feature fusion architecture that aggregates contextual information from different receptive fields, enabling the model to capture both fine-grained local textures and broader semantic structures of surface defects. In addition, we incorporate a progressive dilation strategy, where dilation rates are gradually increased across convolutional layers. This design enhances the model's ability to perceive defects of varying sizes without significantly increasing computational cost or degrading feature resolution. Furthermore, MSFNet-PD employs a lightweight backbone and an efficient fusion mechanism, which collectively contribute to faster inference speed, making the network well-suited for deployment in real-world, high-speed strip steel inspection lines. Extensive experiments conducted on the SD-Saliency-900 dataset demonstrate that our method achieves competitive performance in both detection accuracy and processing speed compared with several recent baselines. The promising results affirm the effectiveness of our approach in practical industrial defect inspection scenarios.