An Enhanced Multi Label Image Classification Framework Using Semi Supervised Deep Learning Techniques

Abstract

The subject of multi-label image classification (MLIC) has attracted a large newlinenumber of researchers in the field of machine learning because of the clarity of the newlineproblem statement and the diversity of possible solutions. Multiple predictions are made newlinefor a single occurrence by multi-label image classifiers. Especially when multiple newlinefeatures and labels are interdependent, the challenge becomes more difficult as the newlinenumber of labels and features grows. Modeling the co-occurrence interdependence with newlinepairwise compatibility probability then inferring the final joint label probability using newlineMarkov random fields is a typical method. In addition, the majority of these approaches newlineeither fail to account for higher-order correlations or make significant computational newlinecomplexity trade-offs in order to simulate more intricate label associations. Another newlinemost challenging task in multi-label image classification is to deal with incomplete, newlinenoisy and missing labels of an image. Multi-label machine learning techniques struggle newlinewhen labels are missing or partial. To address this issue, current methods employ a newlinesingle data representation made up of all features for label classification newline