Wear-resistant steel balls produced by forging often exhibit poor roundness and flash defects, which severely impact their grinding performance. To address this issue, this paper proposes an online visual inspection method for high-temperature wear-resistant balls. By calculating the difference between the maximum and minimum distances from the center of the grinding ball to the contour in the image, roundness is quantitatively represented, allowing for the selection of grinding balls with poor roundness. For flash detection, this paper utilizes a deep learning strategy to effectively identify flash according to certain rules, distinguishing the complex textures of the background area and enabling effective model training. Moreover, capturing the grinding balls at high temperatures using digital filtering imaging techniques effectively removes thermal radiation noise, resulting in clear images of the grinding balls. This paper achieves a 95.3% detection rate of flash using the YOLOv5 instance segmentation model, meeting the technical requirements for online inspection.