非对称图像特征位置误差参数求解校正方法

Correction Method for Solving Position Error Parameters of Asymmetric Image Features

图像辐射强度值受到大气透过率的影响而发生改变,图像特征易发生几何失真,且特征的畸变于其几何位置具有复杂的非线性关系,导致特征位置误差校正难度较大。为此,提出基于深度学习的图像特征位置误差校正方法。将参数代入到图像灰度插值算法中,不断调整参数值,改善图像因非均匀、非对称造成的特征点模糊现象,提取得到图像的特征点位置。基于此,构建AlexNet学习结构,构建待校正误差图像模型和校正模型。利用牛顿迭代法对校正模型中的各项参数求解校正模型,实现对图像特征位置误差的校正。实验测试结果证明,研究方法的应用损失率在实验迭代次数达15次时降至5%以下,均方根误差始终低于0.8pixel,均方误差在实验迭代次数为50次时降至10-4,说明研究方法可精准求出位置误差值和提高图像精度,校正前后图像间特征拟合程度高。

The radiation intensity value of the image changes due to the influence of the atmospheric transmittance.The image features are prone to geometric distortion,and the distortion of the features has a complex nonlinear relationship with their geometric positions,which makes it difficult to correct the error of the feature positions.Therefore,a method of image feature position error correction based on depth learning is proposed.The parameters were substituted into the image gray interpolation algorithm at first.Meanwhile,the parameters were continuously adjusted to improve the fuzzy feature points caused by non-uniformity and asymmetry of image,whereupon the position of feature point was extracted.On this basis,AlexNet learning structure was constructed.After that,a model with the error image to be corrected and a calibration model were constructed.Finally,Newton iteration method was adopted to solve each parameter in the calibration model,thus correcting the position error of image feature.Following conclusions can be drawn from experimental results.The application loss rate of the proposed method can be reduced to less than 5%when the number of iterations is 15,and the root mean square error is always less than 0.8 pixel.The mean square error is reduced to 10-4 when the number of iterations is 50,indicating that the method can calculate the position error accurately and improve the image accuracy.The feature fitting degree between the images before and after correction is high as well.