As the manufacturing industry rapidly advances, the demand for precise workpiece size measurement continues to grow. Efficient and accurate three-dimensional measurement of workpieces is crucial for ensuring processing quality. This paper proposes a detection method for the three-dimensional size of machined workpieces using a robot equipped with a binocular vision system. A flange plate is chosen as a typical detection object, and a visual detection system is developed along with a corresponding workpiece size detection algorithm based on binocular vision. To address the issue of highlight and noise interference in flange images, which leads to pixel value contamination, a gray-level aggregation algorithm is introduced. This algorithm improves the robustness of stereo matching cost calculations by detecting and reconstructing contaminated pixel values. Additionally, to tackle the challenge of large matching errors caused by weak or repeated image features in the flange, a weight adaptive calculation algorithm is proposed to enhance stereo matching accuracy by effectively characterizing pixel features. Building on this, an AD-Census stereo matching optimization algorithm is developed, combining gray-level aggregation and weight adaptive calculation, with its effectiveness validated through flange size detection experiments. Furthermore, by analyzing the transfer process of parallax errors during flange visual inspection, an evaluation model for camera measurement pose is established, allowing the determination of the optimal measurement pose. Flange size detection experiments under different poses confirm the effectiveness of the proposed pose optimization method. The results show that the proposed method significantly improves workpiece size detection accuracy and offers a new technical approach for three-dimensional size measurement of machined workpieces.