Small sample character and freehand sketch recognition method and device

Abstract

The invention provides a small sample character and freehand sketch recognition method and device, and the method is characterized in that the method comprises the following steps: erasing unlabeled source data in a point sequence format according to a fixed erasing proportion, and obtaining augmented network pre-training data; building a BERT augmented network based on a Gaussian mixture model, and training based on augmented network pre-training data and unlabeled source data in a point sequence format to obtain an augmented device; erasing the labeled small sample data in the point sequence format according to each random erasing ratio to obtain erased small sample data; respectively predicting states and coordinates of the erased small sample data by using an augmenter to obtain prediction points, integrating the prediction points with the erased small sample data, and converting by using a neural renderer to obtain bitmap format augmented data; and training a convolutional neural network classifier based on the augmented data in the bitmap format and the labeled small sample data in the bitmap format to obtain a small sample character and freehand sketch recognition model, thereby recognizing the to-be-recognized image to obtain a classification result.

Description

technical field [0001] The invention belongs to the technical field of data recognition, and in particular relates to a small-sample character and hand-drawn sketch recognition method and device. Background technique [0002] Deep learning models have revolutionized visual recognition tasks, but the model performance is largely due to a large number of labeled training sets. However, due to the high cost of data labeling and the scarcity of natural data for certain categories (such as rare species, ancient pictographs, etc.), the deployment of computer vision models in practical tasks is greatly limited. Since humans can effectively learn new visual concepts and recognize new objects from very few labeled examples, the research on few-shot learning is motivated, and its main goal is to use smaller data sets to train robust A better classifier. [0003] In typical small sample learning, the model first learns transferable and general knowledge or representation on a large n...

AI Technical Summary

Problems solved by technology

However, the existing augmentation methods in the field of small samples often also need to label small sample data; the most common data augmentation methods are mostly natural image design (flip, rotation, cropping or color dithering), and the transformation is performed on the image level. Can take into account the diversity of character or sketch local shape

Images