Image classification method and device

Abstract

The invention provides an image classification method and device. The method comprises the steps of classifying a plurality of to-be-labeled images and different target features in a manner of matching the to-be-labeled images with the different target features to obtain a plurality of first categories; classifying the plurality of to-be-labeled images in a clustering mode to obtain a plurality ofsecond categories, and for each first category, determining a second category matched with the first category, taking the to-be-labeled image in the first category, which is the same as the to-be-labeled image in the determined second category, as a to-be-labeled image included in the target category. According to the image classification method and device, the multiple target categories and theto-be-labeled images included in each target category are obtained, so that accurate classification of the multiple to-be-labeled images is achieved, and the problem that too much time cost and laborcost are wasted when fine classification is conducted on the images is avoided.

Description

technical field [0001] The present application relates to the field of image classification, and in particular, to an image classification method and device. Background technique [0002] The difference and difficulty of refined image classification compared with general image classification tasks is that the granularity of the image category in refined image classification is finer, and the visual difference between different subcategories under the same large category is extremely small. [0003] At present, refined images are usually classified according to the amount of supervision information used. Specifically, the refined classification task model can be divided into two types of models: strong supervision and weak supervision. The strong supervision model needs to use a large amount of labeling information. It takes a lot of time and labor costs, and the refined classification is more professional, but it is difficult for most people to distinguish effectively, so th...

AI Technical Summary

Problems solved by technology

[0003] At present, refined images are usually classified according to the amount of supervision information used. Specifically, the refined classification task model can be divided into two types of models: strong supervision and weak supervision. The strong supervision model needs to use a large amount of labeling information. It takes a lot of time and labor costs, and the refined classification is more professional, but it is difficult for most people to distinguish effectively, so the limited cognition level of labelers often leads to unsatisfactory data sets, which will affect the labeling results; and The weakly supervised model can automatically capture local features with less labeling information. Although the labeling cost is reduced, there is always a gap with the strong supervised model in terms of accuracy.

Images