Automated Detection of Defects with Casting DR Image Based on Deep Learning

Abstract

In the production process of castings, due to the constraints of casting technology and production conditions, different types and degrees of defects will inevitably occur inside the casting. The traditional defect detection needs to manually determine the position and type of the defect in the Digital Radiography (DR) image of the casting, which is affected by the subjective initiative of people. For the disadvantages of manual judgment, we propose a method for detecting defects in DR images of castings based on deep learning. Firstly, the image is smoothed by guided filtering, and perform image enhancement on the smoothed image. Subsequently, we propose an improved Cascade Mask R-CNN network based on the Cascade Mask R-CNN to realize the detection and grading of the defects in the DR image of the casting. The experimental results show that the improved Cascade Mask R-CNN network has greatly improved the shrinkage defect recall rate of castings, reduced the missed defect rate, and improved the detection precision. It is proved that the improved Cascade Mask R-CNN network can better realize the detection of casting defects.