Training method of image recognition model, and recognition method and device of text image
By adjusting the model parameters and vector set of the image recognition model, it is possible to achieve both text recognition and structured extraction functions, which solves the problem of accumulated error in the traditional multi-model architecture and improves the recognition effect and training efficiency of the image recognition model.
CN116994260BActive Publication Date: 2026-06-12JD DIGITS HAIYI INFORMATION TECHNOLOGY CO LTD
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JD DIGITS HAIYI INFORMATION TECHNOLOGY CO LTD
- Filing Date
- 2023-08-15
- Publication Date
- 2026-06-12
Smart Images
Figure CN116994260B_ABST
Abstract
The application discloses a kind of training method of image recognition model, text image recognition method and device.The method comprises: input training text image and initial vector set into the initial image recognition model that is not trained completely, obtain the output predicted text recognition data;Based on predicted text recognition data and standard text recognition data, the model parameters of initial image recognition model and initial vector set are adjusted to obtain target image recognition model and reference vector set;Wherein, at least one initial vector combination is contained in initial vector set, initial vector combination includes initial text vector, initial position vector and initial structure vector, predicted text recognition data includes the predicted text content data, predicted text position data and predicted text structure data corresponding to each initial vector combination respectively.The embodiment of the present application solves the problem of cumulative error existing in the traditional multi-model architecture, improves the overall training efficiency and text recognition effect of the model.
Need to check novelty before this filing date? Find Prior Art