【摘要】 关键零件内部复杂结构的精密测量是高端制造领域攻克的难题。当采用工业CT技术实现对象内部结构精密测量时，面临目标图像灰度不均匀性、边缘模糊、伪影等问题。有鉴于此，本文研究了局部能量最小化模型(RSF)的图像分割方法，引入自然梯度和AdamW算法分别提高了RSF模型的收敛速度和参数自适应性。首先，在统计流形上计算自然梯度，提高梯度下降效率和RSF模型收敛速度；其次，采用AdamW算法实现RSF模型的高斯核函数尺度大小自适应控制。与经典RSF模型相比，改进后的RSF模型迭代次数减少了1 353次，迭代次数降低约76.79%,迭代时间减少约43.61%,测针球面半径和航空燃油喷嘴圆柱直径测量误差均较小，既保持了原模型亚像素分割精度，又大幅提高了模型收敛速度和鲁棒性。更多还原

【Abstract】 The internal complex structure precision measurement of key part is a challenge in the field of high quality manufacturing. When the industrial CT technology is used to achieve precise measurement of the internal structure of the object, it faces problems of grayscale inhomogeneity, blurred edges, and artifacts of the target image. In view of these, the local energy minimization model(RSF) image segmentation method is investigated in this article. The natural gradient and AdamW algorithms are used to improve the convergence speed and parameter adaptivity of the RSF model, respectively. First, the approximate natural gradients are computed on the statistical manifold to improve gradient descent efficiency and RSF model convergence speed. Secondly, the AdamW algorithm is utilized to realize the adaptive control of the scale of the Gaussian kernel function of the RSF model. Compared with the classical RSF model, the improved RSF model reduces the number of iterations by 1 353, the number of iterations by about 76.79%, the iteration time by about 43.61%, and the low measurement errors of the probe-radius and the diameter of jet fuel nozzle cylinder, which not only maintains the sub-pixel segmentation accuracy of the original model, but also significantly improves the convergence speed and robustness of the model.更多还原