Study On Class Imbalance Problem In Multi-Lable Image Classification

Logistic regression(LR)is the most commonly used loss function in multi-label image classification,yet it suffers from class imbalance problem caused by the huge difference in quantity between positive and negative samples as well as between different classes.In this work,we first point out that these two imbalances can be quantified by inputting noise samples to the classifier.We find that the predict value of noise samples varies considerably between different classes,even if the noise sample does not contain any valid information for classification.We believe that these differences,which do not depend on the test samples' information,reflect the inference tendency of the classifier and can be used as a basis to rectify the classifier in class imbalance learning.There are two main problems in the traditional class imbalance learning.First,there is only qualitative discussion on the deep mechanism of category imbalance affecting the final prediction,and there is no quantitative analysis.Second,there is no quantitative index to measure the degree of the model affected by the class imbalance.As a result of the above two points,most of the solutions to the class imbalance problem are based on heuristic weighting strategies and lack of systematic theoretical basis.This paper takes image multilabel classification as the benchmark task and gives a detail discussion on the cause of the negative effects of the class imbalance in multi-label classification.This work reveals the relationship between class imbalance andinference tendency,the proposed noise statistics can quantify the degree of imbalance in time with minimal cost.This work shows a new way to solve the problem of class imbalance,by reducing the impact of inference tendency on final prediction.Both proposed extremum shift(ES)and noise constraint(NC)are able to weaken inference tendency during training.Comparative experiments with Resnet on Microsoft COCO,NUS-WIDE and DeepFashion demonstrate the effectiveness of noise statistics and the superiority of our approach over the baseline LR loss and several state-ofthe-art alternatives.