Method and device for establishing multi-attribute recognition model and multi-attribute recognition method

Abstract

Embodiments of the present invention provide a method and device for establishing a multi-attribute recognition model, and a multi-attribute recognition method, wherein the establishment method includes: inputting sample images pre-labeled by multi-attributes into the first model for learning, and obtaining the value of each image in the sample images feature matrix; input the feature matrix of each image to the second model for learning, and obtain the semantic-spatial feature matrix of each image; obtain the first predicted value of each attribute according to the feature matrix, and obtain each attribute according to the semantic-spatial feature matrix The second predictive value of the first predictive value and the second predictive value are weighted and summed to obtain the comprehensive predictive value of each attribute of each image; when the comprehensive predictive value obtained after learning and the label value When the loss of is stable within the preset threshold range, it is determined that a multi-attribute recognition model is learned. The embodiment of the present invention can effectively use the annotation information to obtain the spatial and semantic association of multiple attributes, and the recognition accuracy is high.

Description

technical field [0001] Embodiments of the present invention relate to the technical field of machine learning, and more specifically, relate to a method and device for establishing a multi-attribute recognition model and a multi-attribute recognition method. Background technique [0002] Traditional pedestrian multi-attribute recognition methods include multi-label SVM and Softmax classifier methods, but the accuracy of these methods is not as high as convolutional neural networks. [0003] At present, the methods of using convolutional neural network for multi-attribute recognition mainly include: 1) Using the form of single-attribute multi-model, one model identifies one attribute in a targeted manner, and finally integrates the output results of multiple models to complete multi-attribute recognition ; 2) Adopt the form of multi-label, such as using deep learning frameworks such as MxNet and Pytorch, and directly input multiple labels for learning. During the training pro...

AI Technical Summary

Problems solved by technology

[0004] Using a single-attribute multi-model method to achieve multi-attribute recognition is not efficient; based on the existing deep learning framework, multi-label learning is used to treat all attributes of the entire image equally, ignoring the correlation between attributes. The complex and diverse content in multi-label images makes it difficult to learn effective feature representations and classifiers; dividing images into regions also ignores the correlation between attributes, and the implementation of region proposals is relatively complicated and not very practical

Images