Refinement of CNN Based Multi-label Image Annotation

Abstract

With the mushroom of technology, digital images are increasing rapidly, handling of these images has become an important research issue. Automatic image annotation (AIA) is a method for finding proper labels to an image in order to get a suitable way for searching and indexing the image data. AIA performs an influential role in image retrieval and its management. Exploiting the correlation among labels is a vital task in AIA for solving the semantic gap problems. Finding a contextual correlation among concepts can be helpful further to reduce this gap. To ensure effective capturing of this correlation, this paper presents co-occurrence patterns of labels along with random field methods for improving the performance of AIA. First, the DenseNet201 model is trained as a concept classifier for images and labels associated with images. Based on the training samples and concept vocabulary, co-occurrences of concepts are determined using association rule mining. The conditional random field (CRF) is used for refining the concept predicted by DenseNet201 CNN based on co-occurrence patterns. The experiment is carried out on the LableMe dataset. The performance analysis is carried out using the F1 score, recall and precision. From the obtained results, it is perceived that the proposed approach performs better than the DenseNet201 model.