Detection of circular parts is proposed based on improved Hough transform algorithm. 70%~80% of the edge pixels of the circular parts image are randomly sampled based on the minimum error, which reduced the accumulation of irrelevant data. The clustering algorithm is avoided the appearance of false center, and the roundness is judged by the eligibility of actual circular parts, which give the process. The experimental results show that the parameter value of the improved Hough transform algorithm is closest to the standard value, and the average uncertainty of the circular parts were 0.053 3, which is 35.55%, 32.10%, 26.48%, 18.63%, 8.26% lower than CNN, HT, RHT, PHT and SHT, the measurement accuracy is higher than other algorithms.