A Study On Instance Retrieval Using Deep Learning

The task of instance retrieval is to find the images containing a specific object from a large number of images,which is a hot topic in the field of image retrieval.The instance retrieval is based on the original retrieval task,retrieving a specific object in the image,making the purpose of retrieval more explicit and the applied scene closer to the actual life.This article focuses on the three problems of instance retrieval: 1.Region detection based on deep neural network.Deep neural networks have achieved good accuracy in solving region detection tasks,but often require fine-tuning for datasets in specific problems.For real-time retrieval tasks,we should select the appropriate network structure,without affecting the detection results under the premise of improving search speed;2.Instance feature extraction and encoding algorithm based on the deep neural network.Different from the traditional manual feature extraction method,we design a new retrieval feature by training a deep neural network model from the output of a layer in the network as the original feature;3.The design of the index structure and refining matching results in retrieval system.With the increasing number of images in the database,the time cost is very large if the general retrieval strategy is used.However,the retrieval matching module needs to ensure the retrieval accuracy while ensuring the real-time feedback of the system.The main achievements and innovations of this paper are as follows: 1.For the specific task of instance area detection,we improved the current popular SSD framework,not only achieved high detection accuracy,but also made the retrieval system meet the real-time,but also for retrieval Based on the deep convolutional network model,we propose a feature extraction method based on the convolution activation response distribution,and extract multi-scale features,such as low-level visual features,high-level semantic features at different network layers,So that the instance features used for detection are more capable of representative;3.For the image database of large-scale retrieval,we propose an index data structure suitable for rapid retrieval,which makes it possible to realize the instance retrieval system of large-scale images.On this basis,we propose a reordering algorithm based on region optimization,which makes the retrieval result more accord with the user's expectation.We crawled a large number of natural images from Flickr and other image sites,and for each picture of the instance area of the annotation and classification of object categories,structured stored in the database.We combined with the public dataset and the crawled image data to train the deep neural network model,tested the area of the instance,extracted the feature of the instance,the accuracy and time cost of the retrieval results.The experimental results show that our retrieval framework not only achieved good accuracy on the open dataset,but also achieved real-time response on the large-scale image retrieval task.