Research On Image Retrieval Based On Image Salient Region Features And Deep Learning

Content-based image retrieval can overcome the shortcomings of text-based image retrieval in subjectivity and ambiguity.And the description and feature extraction of image content are the most critical factors which determine content-based image retrieval performance.With the development of visual saliency theory and deep learning techniques,people try to narrow the semantic gap between the visualized feature and image semantic from the perspective of visual saliency and the use of convolution neural network,and there are some fruitful achievements.Based on the analysis of image saliency,this paper presents an image retrieval method based on image salient region features and deep learning.At the same time,we design and develop an image retrieval system to verify the performance of the algorithms in this paper.First of all,in order to accurately extract the image region,this paper designs an automatic image segmentation algorithm based on image saliency map-AutoGrowCut.The saliency map is calculated at multiple scales,we obtain binary image by threshold segmentation,and then use the morphology processing to get labels which accurately indicate salient region and background region,the salient region is extracted using cellular automata algorithm on the basis of labels.AutoGrowCut reduces the manual work and basically achieves artificial segmentation results at the same time.In this paper,we apply the segmentation algorithm to image retrieval,and design the image retrieval model based on salient region features.This model can accurately describe the features of the salient object and improve image retrieval precision.Then,in order to narrow the semantic gap in image retrieval,this paper studies a classification method based on image salient region to obtain the semantic category of query image.We have trained the convolutional neural network classifier based on the image salient regions instead of original images,and obtained the significant semantic category of the image directly,which overcomes the interference of the background information.Images from THUR and ImageNet are used for testing,the results show that this method can further enhance the classification accuracy,and the method has a good robustness for images with complex background.Finally,based on the image retrieval algorithms and models proposed in this paper,we design and develop an image retrieval system.The system includes an off-line image salient region features library establishment module,an off-line convolution neural network classifier training module,an image online category detection module and an image online retrieval module.Experiments show that the image retrieval precision is improved from 68.75% based on original image features to 81.25% based on salient region features.After considering the deep semantic,the precision can be further optimized.This verifies the effectiveness of our image retrieval method based on salient region features and the deep learning.The image retrieval model studied in this paper accords with human visual perception characteristics,which has theoretical and practical value for information extraction and image retrieval.