Blind detection of circular image rotation angle based on ensemble transfer regression and fused HOG

INTRODUCTION: Aiming at the problems of low accuracy in estimating the rotation angle after the rotation of circular image data within a wide range (0°–360°) and difficulty in blind detection without a reference image, a method based on ensemble transfer regression network, fused HOG, and Rotate Loss is adopted to solve such problems. METHODS: The proposed Rotate Loss was combined to solve the angle prediction error, especially the huge error when near 0°. Fused HOG was mainly used to extract directional features. Then, the feature learning was conducted by the ensemble transfer regression model combined with the feature extractor and the ensemble regressors to estimate an exact rotation angle. Based on miniImageNet and Minist, we made the circular random rotation dataset Circular-ImageNet and random rotation dataset Rot-Minist, respectively. RESULTS: Experiments showed that for the proposed evaluation index MSE_Rotate, the best single regressor could be as low as 28.79 on the training set of Circular-ImageNet and 2686.09 on the validation set. For MSE_Rotate, MSE, MAE, and RMSE on the test set were 1,702.4325, 0.0263, 0.0881, and 0.1621, respectively. And under the ensemble transfer regression network, it could continue to decrease by 15%. The mean error rate on Rot-Minist could be just 0.59%, significantly working easier in a wide range than other networks in recent years. Based on the ensemble transfer regression model, we also completed the application of image righting blindly.