Research On Deep Learning And Visual Attention Technology For Accurate Image Understanding

Image comprehension is an important aspect of image processing and its application,such as image recognition,object detection and analyzing the relationship among objects in an image,then forms the semantic understanding of the content in the images.Image comprehension involves many aspects,such as object recognition,scene understanding,object localization,large-scale image retrieval,image segmentation and image annotation so on.This paper is mainly studies the problems of image retrieval and image automatic annotation.In the area of computer vision,image process and deep learning has become a breakthrough in the field of artificial intelligence.Deep learning has a strong power to express and to learn feature of an image,and the rapid increase of computing power provide the condition for us expand deep neural networks s in both depth and width,which can be trained with large-scale data.Both the image retrieval and automatic annotation need to understand the content of an image.To deal with the similarity comparison between two images and image annotation problem,we just need to focus on the specific part of them,rather than the whole image.We propose an new framework based on fine-grained image process to solve the problems of image retrieval and image annotation.Feature representation is important in image retrieval,because it can guarantee the recall rate of the result in image retrieval.In this paper,we propose a new algorithm for image retrieval based on Proposal and deep convolution neural networks to improve the performance of image retrieval.Firstly,an unsupervised proposal method is used to generate the candidate frame which may contain the target,and then we use deep convolution neural networks to extract the regional feature of this image.Finally,the fine-grand features will be used to train the model and learn hash functions.We study image retrieval on three public datasets and compare with the state of art,the results show that the new method can accurately represent the effective information in the image and improve the performance on complex image.In order to solve the problem of automatic image annotation,we propose recurrent neural networks based on visual attention model to label the image one concept once a time.The existing methods,such as using the hand extracted low-level feature and matching them with the existing pattern or using the feature of neural networks s to predict the latent labels of it.We process the local regions of an image so as to recognize them accurately.And we validate the use of the attention with the state of art performance.In this paper,we study the relevant technologies of deep learning,combining with the specific problems in image understanding.We proposed new methods and validate the effectiveness of them in actual situation.