Retrieving images from medical database is termed as medical image retrieval process. Traditional image retrieval is done using various algorithms. There are various limitations in those retrieval process including, manual image annotation, difficulties in handling complex queries, inefficient feature extraction, increase in time requirement which leads to inaccurate results. In this work, these difficulties are rectified by proposing an effective image retrieval method. This work proposes a DCNN based medical image retrieval (CBMIR) framework using contents. Hash coding also incorporated in the proposed work. The unwanted data in the images are removed by using a Gaussian filtering approach in pre-processing stage. From images, deep feature representation is extracted by using a deep Convolutional Neural Network (CNN) model. Distinguished binary codes are generated by designing a new loss function in coding process. The desired binary values are obtained by approximating the real outputs values by adding a regularization term in training process. From trained network, achieve compact binary hash code of query image in retrieval phase. Hash codes of database images are compared with derived hash codes. The experimental results shows that proposed work has better results than the available CNN and hash methods. Feature extraction using a CNN has lot of advantages when compared to traditional methods. Pair of images are used as an input in the adapted Siamese network. The contrasting loss function and weight values are used to make model to learn to make images for particular class based on similar features.
Volume 11 | 11-Special Issue
Pages: 208-214
DOI: 10.5373/JARDCS/V11SP11/20192949