METHOD AND DEVICE FOR TRAINING MULTI-LABEL CLASSIFICATION MODEL

摘要

Provided are a method and device for training a multi-label classification model capable of dynamically learning image features, enabling a feature extraction network to better adapt to task requirements, and providing good multi-label classification performance. The method comprises: determining, from a training dataset, n samples and a label matrix Y_c*n corresponding to the n samples, where an element y_i*j in the label matrix Y_c*n indicates whether the ith sample comprises an object denoted by the jth label, and c indicates the number of labels related to the samples; using a feature extraction network to extract a feature matrix X_d*n of the n samples; using a feature mapping network to acquire a predicted label matrix for the feature matrix X_d*n, where an element in the predicted label matrix indicates a confidence value as to whether the ith sample comprises the object denoted by the jth label; and updating, according to the label matrix Y_c*n and the predicted label matrix, a weight parameter Z and a feature mapping matrix M_c*d to train a multi-label classification model.