Data preprocessing and data enhancement
Data enhancement technology has been proven especially depth training machine learning architecture in favor of a common, whether it is accelerating the convergence process or as a regular item, which also avoids excessive proposed merger enhances the model generalization ability [15].
Enhancement typically includes applying a series of data migration technology in feature space or data space (or both are). Data on Space Applications, the most common enhancement technology, this enhanced technology migration method to get a new sample from the existing data. There are many methods available migration: translation, rotation, distortion, scaling, color space conversion, clipping and the like. The goal of these methods are constructed by generating more samples larger data sets, and to prevent over-fitting the model regularization can also balance the size of each class of the data set, and even manually for the current task or application scenario with a new sample more representative.
Data enhancement is particularly useful for small data sets, but its effectiveness has been proven in long-term use. For example, in [26], there are 1,500 data sets portrait image Design of new dimensions (0.6,0.8,1.2,1.5), four new rotation angle (-45, -22,22,45) , as well as four new gamma changes (0.5,0.8,1.2,1.5) has been enhanced to 19,000 data sets of training images. Through this process, when used to fine-tune the enhanced data sets, which overlap accuracy portraits segmentation system (IOU) increased from 73.09% to 94.20%.