一种基于U-NET多景深图片目标物定位自动聚焦算法

在多景深场景下，已知目标物类型，当目标物位于图像中心位置时，传统的聚焦评价函数曲线灵敏度较低；当目标物偏离中心位置时，聚焦评价函数曲线容易出现局部极大值或无法准确判断出准焦图像，影响自动聚焦系统。针对这两种情况，提出了一种基于U-Net神经网络判断目标物位置，设定对应窗口和评价函数的方法，即当目标物位于图像中心位置时，提出了一种新的聚焦评价函数——SMD-Roberts函数；当目标物不在图像中心位置时，设定对应窗口，选择SML评价函数对图像像质进行评价。实验结果表明，与传统的灰度梯度自动聚焦评价函数和传统的取窗法相比，该方法得到的聚焦评价函数灵敏度最少提高0.0241，耗时最少减少0.0355 s，单峰最少减少1个次峰，有效地解决了多景深场景下，应用聚焦评价函数判断目标物最清晰位置不准确及聚焦评价函数曲线出现双峰的问题，明显地提高了评价函数的无偏性、单峰性以及灵敏度。该方法普适性强，更适用于自动聚焦系统。

An automatic focusing algorithm based on U-Net for target location in multiple depth-of-field scene

Evaluation function of automatic focusing system is the key to evaluate image quality. In multi-depth-of-field scenarios, when the target is located in the center of the image, the sensitivity of the traditional focusing evaluation curve is low; when the target deviates from the center, the focus evaluation function curve is prone to local maximum, which affects the accuracy of the automatic focusing system. In view of these two situations, this paper proposes a method based on U-Net neural network and sets the corresponding window and evaluation function. When the object is located in the center of the image, a new focusing evaluation function, SMD-Roberts function, is proposed. When the target is not in the center of the image, the corresponding window is set for the image and the SML evaluation function is selected to evaluate the image quality. Experimental results show that , compared with traditional focused evaluation function and central window method, this method can effectively solve the problem that the focus evaluation function is not accurate in judging the clearest position of the object and the double peak of the focusing evaluation function curve in multi-depth-of-field scenes and obviously improve the unbiasedness, unimodal and sensitivity of the focused evaluation function. This method has strong universality and is more suitable for focused evaluation system.