CNN Photo Horizon Correction Method based on convolutional neural network and residual network structure

摘要

The present invention relates to a method of measuring an inclination of an image when the given input is not horizontal, and then using the same to straighten the image horizontally. At the end of the network structure, a feature map having a size of 1x1 is output. Having a first pulling layer; A second pulling layer having a feature map of size 2 × 2 as an output; And an angle predictor that combines the feature vector obtained from the first pooling layer and the feature vector obtained from the second pooling layer and uses the resultant prediction as the final result prediction. Generating a training dataset for learning the training; (b) the angle measurement network with different learning settings Optimal parameters Learning; And (c) the angle measurement network with respect to an input image. Measuring an inclined angle using the same, rotating the image in the opposite direction by the measured inclined angle, and cropping the empty pixel area.