Self-supervised learning model training method and device based on relational reasoning

Abstract

The invention provides a self-supervised learning model training method and device based on relational reasoning. The method comprises obtaining different local observation images corresponding to theimages through different geometric transformation operations; extracting local features corresponding to the corresponding images; fusing the local features to obtain global features of the corresponding images; predicting a corresponding prediction geometric transformation operation between the local feature and the global feature; according to the difference between the prediction geometric transformation operation and the actual geometric transformation operation; constructing a loss function of the learning model; determining target parameters of the learning model through iteration of the loss function; using the prediction geometric transformation operation as a supervision signal to train a learning model. The relationship of the preset auxiliary task is established between the global feature and the local feature, so that the feature obtained by model learning can focus on capture of semantic information of the visual object, the influence of the preset auxiliary task on feature learning is reduced, and migration to the target task is easy.

Description

technical field [0001] The invention relates to the technical field of computer vision and machine learning, in particular to a self-supervised learning model training method and device based on relational reasoning. Background technique [0002] The emergence of large-scale annotated datasets is one of the key factors for the great success of deep learning in the field of computer vision. However, there is a major problem with supervised learning: it relies too much on large-scale datasets, and the collection of datasets and manual data annotation require a lot of human costs. Therefore, self-supervised learning methods have recently received extensive attention in the industry. Self-supervised learning methods learn and generate semantic label information of visual features by mining the properties of data. [0003] In related technologies, self-supervised learning aims to learn distinguishable visual features by designing auxiliary tasks, so that target labels can be fre...

AI Technical Summary

Problems solved by technology

However, there is a major problem with supervised learning: it relies too much on large-scale datasets, and the collection and manual data annotation of datasets requires a lot of labor costs

However, the training of the model relies on minimizing the learning objective related to the preset task, so the learned visual representation not only contains the visual semantic information of the input image, but also contains the knowledge related to the auxiliary task, which makes representation learning and the designed auxiliary task The relationship between them is too close, that is to say, the current preset auxiliary tasks have a greater impact on representation learning, hindering the transfer of learned representations to other target tasks, and may lead to performance degradation.

Images