Two-Stage Segmentation Method for Quantifying Composite Micro-Cracks Based on Deformable Convolution and U-Net

Under extreme working conditions, composite materials must endure significant temperature gradients. These gradients can lead to micro-cracking, delamination, and other forms of damage, potentially resulting in leaks and catastrophic failures. Traditional edge segmentation methods struggle to accurately extract crack features due to their irregular geometries. Additionally, cracked images often contain noise and artifacts, and small cracks mixed with complex backgrounds can result in false edges when using deep learning methods directly. This paper proposes a two-stage image segmentation method based on deformable convolution and U-Net. In Stage I, a vertical neighborhood local enhancement algorithm is introduced. A selective search algorithm identifies regions of change, followed by a Gabor filter to enhance longitudinal cracks and generate candidate frames. The algorithm accurately locates the target crack image within the candidate frame. In Stage II, deformable convolution is introduced for feature extraction enhancement. Deep deformable convolution is integrated into the U-Net network to improve adaptability and capture image details. The results demonstrate that our two-stage image segmentation method accurately locates and segments microcracks in composite materials, achieving excellent mean Intersection over Union scores.