Deep learning feature map for content based image retrieval system for remote sensing application

• DOI: 10.1504/ijim.2016.079113
• Publication Date: 2016-09-14T11:30:05Z
• Authors: Swati Jain; Tanish Zaveri; Kinjal Prajapati; Shailee Patel
• Keywords: 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology

This paper proposes a model for content based image retrieval system (CBIR), in which handcrafted feature set is replaced with feature set learnt from deep learning, convolutional neural network (CNN) for image retrieval. Feature map obtained from CNN is of high dimension, which makes the matching process expensive in terms of time and computation. Hence to recapitulate information in smaller dimension statistical values of the feature maps are calculated. Statistical values like entropy and contrast are taken as characteristic value of each feature map, and these values are used as features for similarity measure in CBIR system. The retrieval performance are compared, when feature map from deep neural net is considered and when statistical values of the feature map are used. The performance parameter considered are normalised rank and number of relevant images retrieved. The proposed approach is experimented with UC Merced Landuse Landcover Dataset and the results obtained establishes that statistical features give better results.