DISTILLATION OF NEURAL NETWORK MODELS FOR DETECTION AND DESCRIPTION OF IMAGE KEY POINTS

• Main Authors: Artem V. Yashchenko, Anatoly V. Belikov, Maxim V. Peterson, Alexey S. Potapov
• Format: Article
• Language: English
• Published: Saint Petersburg National Research University of Information Technologies, Mechanics and Optics (ITMO University) 2020-07-01
• Series: Naučno-tehničeskij Vestnik Informacionnyh Tehnologij, Mehaniki i Optiki
• Subjects: deep learning; keypoint detection
• Online Access: https://ntv.ifmo.ru/file/article/19639.pdf
• ISSN: 2226-1494; 2500-0373

Subject of Research. Image matching and classification methods, as well as synchronous location and mapping, are widely used on embedded and mobile devices. Their most resource-intensive part is the detection and description of the image key points. In case of classical methods for detection and description of key points they can be executed in real time on mobile devices but for modern neural network methods with better quality, such approach is difficult due to trading off performance. Thus, the issue of speeding for neural network models applied for the detection and description of key points is currently topical. The subject of research is distillation as one of the methods for reducing neural network models. The aim of the study is to obtain more compact model for detection and description of key points and a description of the procedure for this model design. Method. We proposed a method for pairing the original and more compact new model for its subsequent training on the output values of the original model. In this regard, the new model is learned to reconstruct the output of the original model without using image labels. Both networks accept identical images as input. Main Results. Neural network distillation method for detection and description of key points is tested. The objective function and training parameters that provide the best results in the framework of the study are proposed. A new data set is created for testing key point detection methods, and a new quality indicator of the allocated key points and their corresponding local features is added. New model training in the way suggested with the same number of parameters, shows greater accuracy in key points compared to the original model. A new model with a significantly smaller number of parameters shows the accuracy of point matching close to the accuracy of the original model. Practical Relevance. More compact model for detection and description of image key points is created applying the proposed method. The model is applicable on embedded and mobile devices for synchronous location and mapping. Such model application can also increase the service efficiency of the image search on the server side.