الفهرس 
AbstractOnly 14 pages are availabe for public view
 |  | from | 120 |  |  |
| | | from | 120 | | |
Abstract
Image Ranking is a process of ordering images for search task by comparing them to know if they share the same embeddings. Deep learning has become a breakthrough topic of increasing researchers ’ concern. Deep neural network a groundbreaking tool of deep learning for processing staggering quantities of data, helping to realize a generalization and obtimization for ranking system. In this work we propose an image ranking system that handles the search results based on detection parameters. We hyper-parameterized the detection model before training for increasing the flexibelity. We utilized a deep network and pyramid network for multi-scale features extraction. The detection task is accomplished repeatedly, each time updating it a little to improve the detection result. The significant task of the detection learning model is converged on generating the best learning parameters, which will be used for the ranking model. Transfer learning method is an elegant solution for retrieval task that focuses on inhibiting the learning time of a deep ranking model. The learning parameters are used as a hyper-parameters for the ranking model that have enormous leverage in the retrieval results. We employe triplet ranking loss to make sure that two image example with the same label have their embeddings close together. The Extensive experiments show that the proposed ranking model has significant generalization ability with the accuracy as 86.8% on the VOC 07-12 20K dataset. The proposed ranking model outperformed the model based on hand-crafted visual features, deep classification model, and learning a fine-grained image similarity model based on Alex Net.