【摘要】 当前的畸变矫正技术都是在有参照物的基础上进行的，比如直线、人脸等，应用范围有限。一些深度学习方法也只聚焦于单一的畸变矫正问题，其矫正效果不够好。为了有效解决生活中广泛存在的多种畸变问题，文中提出一种基于RepVGG网络架构的图像混合畸变矫正方法。该方法对于提取到的特征，分别采用预测模型和分类模型获取预测数据和畸变类型数据，通过这两组数据训练出畸变矫正模型，实现畸变图像的预测、分类以及矫正过程。网络还引入空间注意力机制来关注变形严重区域。同时，为了降低模型训练复杂度，引入线性缩放层代替RepVGG块中的非线性结构，使训练时可以合并块中的操作。再通过压缩块的方式将多分支结构转换为单个卷积层，既加速训练过程，同时又保持了多分支结构的表达优势。通过实验证明该模型在6种类型的畸变图像上都有出色的矫正能力。与现有算法相比，该算法的矫正率均可以达到98%以上，模型训练速度提升了约1.53倍，该方法在图像畸变矫正领域具有一定的适用性。更多还原

【Abstract】 Current distortion correction techniques are based on the availability of references, for instance, straight lines and faces, and they have a limited range of applications. Some deep learning methods also focus only on a single distortion correction, and the correction effect is not good enough. Therefore, an image hybrid distortion correction method based on RepVGG network architecture is proposed to effectively eliminate multiple distortions widely existed in life. In the method, for the extracted features, a prediction model and a classification model are used to obtain the prediction data and the distortion type data, respectively, and a distortion correction model is trained based on the two sets of data, so as to realize the process of prediction, classification, and correction of the distorted images. The spatial attention mechanism is also introduced into the network to focus on the region of severe distortion. The nonlinear structure in the RepVGG block is replaced with a linear scaling layer, which allows the operations in the block to be merged during training, thus reducing the model training complexity. Then the multi-branch structure is converted into a single convolutional layer by compressing blocks, which not only accelerates the training process, but also maintains the expression advantage of the multi-branch structure. The model is experimentally demonstrated to have excellent correction ability on six types of distorted images. In comparison with the existing algorithms, the correction rate of the proposed algorithm can reach above 98%, and the training speed of the model is improved by about 1.53 times. To sum up, this method has a certain application in the field of image distortion correction.更多还原