Image retrieval with hand drawn draft based on saliency detection

Abstract

An image retrieval algorithm with hand drawn draft based on saliency detection was proposed and improved. First, salient regions were obtained by an image segmentation method, next region color sparse histograms were computed, and then the histogram was used to compute each region color contrast. The saliency map was improved by an enhancement and suppression method. Salient object was obtained by eliminating low contrast regions and was divided into blocks equally. The average gray value of each block and spatial relationships were feature data for searching. Search process was filling hand draw contour and extracting the feature data and comparing the feature data on disk, then the difference was combined with the spatial weights. Finally the results are sorted, and the pictures are similar if the difference is small. This method can extract good salient objects and obtain accurate retrieval results. Introduction With the development of science and the living demand growth, it becomes more convenient to create massive image data. Pictures contain a large amount of information which is various and complete. Image retrieval is the premise of effective use of image data. Image retrieval on smart devices is mainly based on text, such as name, keywords, description text. This approach cannot express the complete information of the image . Hand-drawn sketches image retrieval is more imaginative and creative than text mode on smart devices, and it has a special expression way to meet human’s demand. Hand-drawn sketches using Fourier shape is more efficient than the stroke description way and insensitive about input sequence, so it can obtain better retrieval results . Retrieval method based on structure diagram reduces the dimension of the image contour effectively with less image shape losing, and provides rotation, translation and scaling invariance. The essence of hand-drawn sketches image retrieval is a kind of content based image retrieval. Existing content based image retrieval methods express the image semantics mostly through the extraction of image low-level feature (color, texture and shape) , and don’t allow computers to understand images in human’s cognitive way. There are mainly two kinds of the latest research, one is the use of the weighted sum approach to comprehensive utilization of various features [5, 6, , another is to construct mixed feature [8, 9, 10, 11, . Color image segmentation with improved watershed algorithm calculates saliency of each sub region with the regional spatial attention model. This method tends to comprehend salient object as a whole, so it causes loss of edge details .If a color distributes widely or dispersedly in an image, it’s unlikely included in the salient object. The specific color’s global spatial distribution can be used to describe space saliency of the salient object. The preliminary classification of colors in the image through the Kmeans clustering algorithm is performed, then model color probability distribution using multivariate normal distribution . Existing search method based on content can overcome some shortcomings of text based search, but there are still problems such as high computational complexity, large storage quantity is large, low efficiency. This paper presents an algorithm which first gets a salient region and then extracts the salient region’s similar characteristics. Image semantic contents focus on salient region, so it can be used to represent the image in image retrieval. International Conference on Information Sciences, Machinery, Materials and Energy (ICISMME 2015) © 2015. The authors Published by Atlantis Press 1085 Saliency map based on image segmentation First, the original image is segmented into regions. Then, each region’s color histogram is calculated using algorithm mentioned in literature[4], the RGB component pixel color in each region of the original image were quantified for 12 values. Next, color histogram of each region is calculated. Colors are ranked according to frequency of occurrence. Low-frequency colors are replaced by similar high-frequency colors, and colors which frequency is 0 are removed, we obtain sparse histogram lastly. So images can keep basically the same visual effect with the original image, but greatly reduce the number of colors. Then each color histogram is converted to Lab space, and color distances among different regions are calculated. The formula of color distance between regions m r and n r is as following: