Deep Learning Approach for Image Retrieval
In the Content-Based Image Retrieval (CBIR) domain, several traditional techniques are proposed in the area of feature extraction from images. However, these methods are focused on the extraction of images’ visual content and were not free of defects in terms of recognition. In the last few years, different works of literature have demonstrated that Convolutional Neural Network (CNN) achieves good results in image retrieval. It has been conceived as a performing method to provide effective descriptors for image retrieval compared to traditional methods for feature extraction. In this paper, a hybrid feature extraction method, which benefits from CNN advantages, is presented. The proposed approach combines two pre-trained models of CNN (Inception-v3 and MobileNet) to extract features and uses the K-means method to improve the performance. Using Wang, Oxford5K and Holidays databases, the proposed hybrid model enhanced the retrieval results in comparison to other CBIR approaches with an average precision score of 82% to 94%.