【摘要】 目前人脸美丽的深度学习回归预测算法都是使用的同一种最简单的回归方式,该种回归方式训练过程中训练集验证集上损失值波动比较大、且它的输出没有上下界限制,不太合理,针对这些问题提出一种新的回归预测层设计方法。该回归层先使用Softmax函数归一化将人脸属于各个美丽级别的概率投影到多个节点上,再对美丽分数求解数学期望,投影到预测分数节点,最终使用均方误差函数对网络进行目标函数优化。实验证明该方法明显优于目前的回归方法,具有训练集误差和验证集误差波动小、输出范围合理、预测精度更高等优点。更多还原

【Abstract】 The deep learning regression prediction algorithm with face beauty is the same simple regression method at present. In this kind of regression method, the loss value of the training set and validation set fluctuates greatly, and its output has no upper and lower bounds, so it’s not so reasonable, proposes a new regression prediction layer design method for these problems. The regression layer first uses the Softmax function to normalize the probability that the face belongs to each beautiful level to be projected onto multiple nodes, then solves the mathematical expectation of the beautiful score, projects it to the predicted score node, and finally uses the mean square error function to target the Function optimization. The experiment proves that the method is obviously superior to the current regression method, and has the advantages of small error fluctuation of the training set and validation set, reasonable output range and higher prediction accuracy.更多还原