Research On Image Representation Method Based On Object-level Deep Feature Aggregation

In the era of big data,compared to the text,the image contains a wealth of content information and has a more intuitive way of expression,which make the image gradually become an indispensable part of big data.But the image data is more complex than the text and audio data.So how to make the computer organize and utilize these image data more effectively becomes a research hotspot.With the development of hardware equipment and the emergence of big data,CNN(Convolution Neural Network)has achieved great success in the field of computer vision.Compared with traditional hand-crafted visual features,CNN can automatically learn the features from big data,and can learn more targeted features from training data for different applications.While the early computer vision algorithm is limited by the hand-crafted visual feature,there are still some ideas that are worth learning from.In addition,the image is usually composed of multiple objects.The understanding of these objects is the key to analyze the image information.Therefore,by combining the deep learning and the concept of feature aggregation,this paper proposes two object-based deep feature aggregation algorithms to generate specific image representation,which are used for content-based image retrieval and fine-grained classification,respectively.In this paper,we first propose an object-based deep feature aggregation method,which represents an image through aggregation of the deep features describing the underlying objects.This method generates an image representation with high robustness to spatial layout of objects in the scene and high invariance to general geometric transformations,which is suitable for content-based image retrieval applications.Then,we propose another object-based deep feature aggregation method,which filters and fuses multi-scale object parts under the condition of only class labels.This method generates an image representation with high discrimination,which is suitable for fine-grained classification tasks.Finally,we verify the effectiveness of the two methods proposed in this paper by comparing them with the state-of-the-art methods on a series of benchmark databases.