Feature Extraction And Distance Measure Prediction for CBIR of Satellite Images

Abstract

CBIR systems are into existence for few decades now. Still the researchers are thriving to gain on the performance based on retrieval results. Feature extraction being very important step should be representing the images appropriately and accurately. Several features sets has been flooded in these years. One more aspects which contribute majorly on the retrieval results is matching algorithm. It is observed that different feature set shows varying result with different matching algorithm. Different matching algorithm has shown different behaviour in terms of retrieval performance for different features. Hence we cannot generalize and conclude a particular matching algorithm for a feature set or for a class of image. We present here a neural network that is trained to predict and pick the best matching algorithm for a specific instance. The above system is experimented with the UC Merced LULC dataset and it was observed that the prediction system picks the best performing algorithm approximately 90\% of the time. The features sets used are Local Binary Pattern (LBP), Local Tetra Pattern(LTrP), Circular Covariance Histogram (CCH), Rotation Invariant Transform (RIT) tested using the different distance matching algorithm such as Euclidean distance, Manhattan distance, Chi-Square distance and histogram intersection distance.