Transferable image recognition method and system based on discrimination confidence level

Abstract

The invention discloses a transferable image recognition method and system based on differentiating confidence levels. First, the source domain data is used to train to obtain a source domain pre-training model, and the parameters obtained by using the source domain model training are used as feature extraction parameters of the target domain model. and classification parameters, so that the target domain model selects pseudo-labeled credible samples from the target domain data based on the training parameters of the source domain model, and uses the selected credible samples to assign pseudo-labels and weights to untrustworthy samples, effectively reducing the The uncertainty of the pseudo-labels of all target domain images; finally, the target domain model is trained and optimized by the target domain data with pseudo-labels and the source domain data, so that the target image recognition performance of the final target domain model has been greatly improved. , can carry out rapid migration and effective image recognition work; and effectively reduce the labeling of target image recognition, greatly reducing manpower and material resources.

Description

technical field [0001] The present invention relates to the technical field of image classification and recognition, in particular to a method and system for transferable image recognition based on differentiating confidence levels. Background technique [0002] Transferable image recognition refers to the technique of using labeled images with similar but different distributions to guide the current unlabeled image for accurate recognition when performing image recognition. In the era of big data, it has become a benign development trend to analyze the value information hidden in data to guide people's life and production. But in real-world scenarios, it is very easy to collect a large amount of unlabeled data, and accurate manual labeling on certain tasks is very time-consuming and labor-intensive, such as the accurate labeling of large-scale sensor images. Under this limitation, we can use the existing annotated images to guide the current image recognition task by using...

AI Technical Summary

Problems solved by technology

However, due to the differences in the internal structures of sensors A and B, there are differences in the distribution of the image data collected by the two, so how to implement the image collected by sensor B (generally called the target domain) in the case of differences in image distribution? image) for accurate recognition is a difficult point in the current transferable image recognition problem

Traditional method: Accurately label the data collected by the sensor, retrain a model, and use the model for image recognition tasks, but this process produces expensive and labor-intensive waste, and in the context of big data, all collected data must be accurately Manual labeling is extremely unrealistic

Images