"Content based image retrieval system by merging multi convolutional neural networks"

Abstract

Content based image retrieval (CBIR) systems use low level image representations to measure image similarity and fetch relevant images. Color, texture and shape properties are considered as important handcrafted features in traditional CBIR systems. The success of such a CBIR system depends on the choice of the handcrafted features being used. To use relevant handcrafted features, one needs to have a good knowledge of the domain where CBIR is being applied. Usage of inappropriate handcrafted features may widen the semantic gap and can to lead to poor retrieval results. Hence it is very important to extract features which are independent of prior domain understanding. In addition, it is beneficial if the features can be learnt automatically from an input image. Machine learning methods can be used for learning valuable representations from input image data. In machine learning area, Convolution neural networks (CNN) are able to create important expressive features from a given image data. Hence CNNs are well suited for image processing applications like classification, object recognition and clustering etc. Very large datasets, huge computing resources and processing time are required to train a deep CNN model effectively. There are many deep CNNs available which are pre-trained on massive datasets and distributed for public use. The knowledge learnt from these pre-trained deep CNN models can be applied to address image processing issues in new domains. In this thesis, we have created a framework to leverage VGG16, ResNet50 and InceptionV3 deep CNN models for extracting important features and use these features for image retrieval task. We apply this frame work for an interesting problem, to retrieve images from wang dataset. Experimental results indicate that the features which are extracted from pre-trained deep CNN model perform better than handcrafted features when used for image retrieval application.