Micro-defect Varifocal Network: Channel attention and spatial feature fusion for turbine blade surface micro-defect detection

Micro-defects on the surface of turbine blades can lead to aviation accidents, thus it is important to thoroughly detect. To realize micro-defect detection, we propose a small object detection model from three aspects: multiscale representation, contextual information and region proposal. The proposed Spatial Feature Pyramid Network (SFPN) achieves feature fusion by integrating cross-scale information and representing the fusion results in a multiscale method. To incorporate contextual information, we introduce a Spatial Attention Residual (SAR) module into SFPN for acquiring global spatial information. Furthermore, to capture global channel information and additional contextual information, we design a Selective Multiple Kernel Networks (SMK) module. For region proposal, inspired by the effective screening strategy employed by Varifocal Network (VFNet), we further optimize our model's screening strategy for positive sample region proposals. Based on VFNet, a Micro-defect Varifocal Network (MVFNet) is designed by integrating SFPN and SMK. MVFNet is validated on the NEU-DET dataset with AP reaching 45.8%, exceeding VFNet by 7.7%. The MVFNet achieves an AP of 73.5% in our self-made Turbine Blade Defect (TBD) dataset, while maintaining a remarkable processing speed of 12 images per second at a resolution of 2448 × 2048 pixels. Furthermore, the detection model has been successfully implemented in a real-time automatic turbine blade detection system.