MDD-Net: A novel defect detection model of material microscope image

The defects in the microscopic image of the material have an important influence on the macroscopic properties of the material. Statistics of the number and size of defects in material microscopic images is an essential task to quantitatively study the relationship between micro-structure and macroscopic properties of materials. This task is repetitive and onerous for researchers. Replacing the manpower in this task with computers is helpful to reduce the burden of researchers and beneficial to the development of material science. In this work, a novel defect detection model of material microscope image is proposed, which is called Material Defect Detection Network(MDD-Net). This paper introduced dilated convolution and tailored the Resnet aiming at the small and medium objects dominating problem, introduced deformable convolution aiming at the irregular shape problem, and applied focal loss solving imbalance between foreground and background. The correctness and feasibility of the above ideas are verified through experiments. With the ideas, MDD-Net achieve 17.17% higher precision and 7.93% higher recall compared with baseline.