Jiahang JIANG

Random smoothing data augmentation is a unique form of regularization that can prevent overfitting by introducing noise to the input data, encouraging the model to learn more generalized features. Despite its success in various applications, there has been a lack of systematic study on the regularization ability of random smoothing. In this paper, we aim to bridge this gap by presenting a framework for random smoothing regularization that can adaptively and effectively learn a wide range of ground truth functions belonging to the classical Sobolev spaces.

According to the idea of pixel classification, a new algorithm for fovea detection based on feature extraction was proposed. Six features extracted from spectral domain optical coherence tomography imaging datasets, which can represent the morphological characteristics of fovea, were used to train the random forest classifier to segment the retinal macular area. Then we took the geometric center of the macular area as the final foveal detection result. The experiment was validated by a five-fold cross validation in a retinal image database containing four different kinds of diseases. The distance error between our automatic detection results and the manual results by doctors is 248.9±206.2μm, and the positioning accuracy (the proportion of cubes with a deviation of less than 750μm) is 96.8%. The experimental results indicate that the proposed algorithm has good robustness and can be used to locate the foveal position accurately, which provides a reliable basis for clinical diagnosis.