Efficient deep learning approach for multi-label semantic scene classification

Abstract

Recognition of semantic scene is a tedious task in which a single scene image is correlated with multiple classes. Also, the semantic scene classification is represented as multi instance multi label (MIML) classification, a part of multi label (ML) classification. Identifying the complex semantic relationship between the classes is the major issue in the traditional Machine Learning Technique (MLT) for MIML learning. Therefore, an efficient Deep Learning framework, Convolutional Neural Network (CNN) with a Gaussian blur filter is proposed. The proposed framework supports to identify the complex class correlations between various classes. Experimental results show that the proposed CNN for ML learning scene classification achieves a better predictive performance of 90% and a hamming loss of 0.13 when compared with the existing Machine Learning Techniques (MLT).