Training method and device for a picture recognition model

Abstract

The invention aims to provide a training method and device for a picture recognition model, and provides a method for quantifying the degree of suitability of a certain subspace to picture data, so that the fitting capability of each channel of high-dimensional features to the data is quantified, the operability to multiple high-dimensional channels is improved, and the accuracy of the picture recognition model is improved. The invention becomes a theoretical basis of data selection subspace. The utilization rate of the high-dimensional space is fully mined, the characteristics of the high-dimensional space are respectively focused on certain data, the complementary capability of the plurality of subspaces is promoted, and the mutual decoupling of the subspaces is successfully realized toa certain extent, so that the training difficulty is reduced, the convergence rate is improved, and the final performance is improved.

Description

technical field [0001] The invention relates to the field of computers, in particular to a training method and equipment for a picture recognition model. Background technique [0002] The purpose of deep metric learning (deep metric learning) is to learn a metric space, so that the distance (metric) of semantically similar objects in the embedding space is close enough, and the distance (metric) of semantically dissimilar objects in the feature space far enough. This is also the common goal of most computer vision tasks, understanding objects semantically. Most of the current methods are to learn a unified metric space for all data. Using triplet or binary loss function constraints, the Euclidean distance (L2 distance) of high-dimensional features of objects with the same semantics is as close as possible, and the L2 distance of different objects is as far as possible. [0003] Mostly because the variance of the data is relatively large and the categories are huge, a unif...

AI Technical Summary

Problems solved by technology

[0003] Mostly because the variance of the data is relatively large and the categories are huge, a unified metric space will increase the difficulty of training. For example, the categories of animals and furniture are semantically different, so forcibly using the same metric space to measure them will be difficult. Increase the difficulty of training

And cannot make full use of the huge high-dimensional space

Images