Processing method and device for image recognition model

Abstract

The embodiment of the invention provides a processing method and device for an image recognition model. Through the method and device provided by the embodiment of the invention, for the image recognition model pre-trained by using a sample set, a first image under a relatively small first resolution is obtained based on a first sample image in the sample set, then, for the first image, a first adversarial sample image is generated by using a pre-trained image recognition model, and then the resolution of the first adversarial sample image is improved to obtain a second image under a larger second resolution, and the pre-trained image recognition model is further trained by using the second image and the first recognition tag corresponding to the first sample image as a first correction sample, so that the corrected image recognition model is used for image recognition. According to the embodiment, the model vulnerability is found through the low-resolution image, and the model vulnerability is repaired through the high-resolution image, so that the robustness of the image recognition model is improved.

Description

technical field [0001] One or more embodiments of this specification relate to the field of computer technology, and in particular to a processing method and device for an image recognition model. Background technique [0002] With the development of machine learning technology, deep learning has made great breakthroughs in tasks such as image classification, image detection, and face recognition, and even in some tasks, the recognition ability of the model exceeds that of human beings. However, studies have shown that the recognition ability of cutting-edge classification models may be affected under the disturbance of perturbation. In other words, when there are distracting factors in the picture, the image recognition results of machine learning may be wrong. For example, the ImageNet C data set shows that image classifiers may have differences in identifying disturbed noise samples with rain, snow, fog and other weather, resulting in robustness problems for image recogn...

AI Technical Summary

Problems solved by technology

For example, the ImageNet C data set shows that image classifiers may have differences in identifying disturbed noise samples with rain, snow, fog, etc., resulting in robustness problems for image recognition models.

Model robustness issues may lead to safety issues or asset losses. For example, errors in digital recognition by the OCR model may lead to entry errors in the financial system; errors in the recognition of road signs by the autopilot system may execute wrong instructions, resulting in irreparable traffic accident, wait

Images