Sorting task is one of the main activities in manufacturing. Traditional industrial sorting technology is laborious, time-consuming and inefficient, and it is difficult to meet the needs of automated long-term operations. Therefore, this paper designed a vision-based workpiece recognition system for intelligent manufacturing, which applied deep learning methods to realize the recognition and localization of workpieces to drive the robotic arm to sort multiple types of workpieces. In this paper, the transfer learning method was used to train the enhanced data to recognize new images. Filtering methods such as Gaussian, Bilateral, and morphological transformation methods such as expansion, erosion, and opening operations was applied to achieve image denoising and distortion elimination. Finally, through the feature matrix calculated by processed image data, the information such as the centroid position and the deflection angle can be obtained, which lay the foundation for accurate localization and rapid sorting of workpieces.