Unsupervised Content Based Image Retrieval Using Pre-Trained CNN and PCNN Features Extractors

Abstract

Content-based image retrieval (CBIR) is a popular approach for searching and retrieving digital images from labeled or unlabeled image collections using image content features. The most important parts of the CBIR system are the computational complexity and the retrieval accuracy. Many studies have been conducted to increase the accuracy of image retrieval systems. In this paper, we propose a new fully unsupervised content-based image retrieval (CBIR) technique to increase the efficacy of image retrieval using a clustering approach on unlabeled image dataset. The proposed method combines features extracted from both pre-trained AlexNet model and pulse-coupled neural networks (PCNN) to extract high and low level features. Then principal component analysis (PCA) is performed on AlexNet's features and these combinations are fed to the K-means algorithm after normalization process. Then Euclidean distance is used to measure the similarity between query and stored images within the same cluster. Finally top similar images are ranked and retrieved. Experimental results on the benchmark Corel-1k and Corel-10k datasets show that the proposed method achieves high precision values of 92.8% and 64.8% respectively, on the top 20 retrieval levels compared to other methods.