Publication

Abstract : The use of Convolutional neural networks (CNN) with deep learning performed an excellent performance in many applications of image processing. The use of CNN based techniques to extract image features from the final layer and the use of a single CNN structure may be used for identifying similar images. Learning feature extraction and effective similarity comparison comprises the Content-Based Image Retrieval (CBIR). In CBIR feature extraction, as well as similarity measures, play a vital role. The experiments are carried out in two datasets such as UC Merced Land Use Dataset and SceneSat Dataset. By using a pre-trained model that is trained on millions of images and is fine-tuned for the retrieval task. Pre-trained CNN models are used for generating feature descriptors of images for the retrieval process. This method deals with the feature extraction from the two fully connected layers, which is present in the VGG-16 network by using transfer learning and retrieval of feature vectors using various similarity measures. The proposed architecture demonstrates an outstanding performance in extracting the features as well as learning features without a prior knowledge about the images. By using various performance metrics, the results are evaluated and performance comparison was done. Cosine Similarity and Euclidean Distance performs better in both Fully connected layers.