Researchers Present Multi-Scale Fusion Network for Real-Time Steel Defect Detection

A research team has introduced a convolutional neural network framework named Multi-Scale Fusion Convolution Network with Progressive Dilation, or MSFNet-PD, for detecting surface defects on strip steel. The framework targets real-time inspection needs in industrial manufacturing where high-resolution images and varied defect patterns must be processed quickly.

The network uses a multi-scale feature fusion architecture that gathers contextual information from different receptive fields. It also applies a progressive dilation strategy that increases dilation rates across layers to capture defects of varying sizes while keeping computational costs low.

MSFNet-PD employs a lightweight backbone and an efficient fusion mechanism to support faster inference speeds. These elements allow the model to handle both fine-grained local textures and broader semantic structures without heavy attention modules or complex upsampling.

The approach addresses two common limitations in existing salient object detection models: heavy backbones that slow inference and fixed dilation strategies that limit scale adaptability. By gradually expanding receptive fields, the network maintains spatial resolution while expanding context coverage.

Experiments were conducted on the SD-Saliency-900 dataset.

Results showed that MSFNet-PD achieved competitive performance in both detection accuracy and processing speed compared with several recent baseline models. The authors state that the method supports real-time deployment on high-speed strip steel inspection lines. They note that the framework balances accuracy and efficiency without requiring additional heavy computational resources.