The forged wear-resistant steel balls in production often exhibit poor roundness and burrs, which significantly affect their grinding performance. To solve this problem, an industrial vision inspection method and system is proposed. Roundness of the ball is calculated by the maximum difference between the distances from the ball""s center to its contour. For the default of burr detection, a deep learning detection model is employed. Certain rules to distinguish burrs from the complex textures of the background regions are regulated, which enables the model to be trained effectively. Through analysis of burr features, it is found that burrs often appear as protrusions at the contours and exhibit stripe patterns in terms of brightness and slope. Additionally, capturing images of the high-temperature steel balls using digital filtering imaging effectively removes thermal radiation noise and obtains clear ball images. These images are applied to the YOLOv5 instance segmentation model, resulting in a burr detection rate of 95.3%.