International Journal of Modern Trends in Engineering and Research www.ijmter.com e-ISSN No.:2349-9745, Date: 28-30 April, 2016 @IJMTER-2016, All rights Reserved A Novel Approach for Sketch Based Image Retrieval with Descriptor Dipika R. Birari 1 ,Prof. J. V. Shinde 2 1 Department of Computer Engineering, Late G. N. Sapkal COE, Nashik, dipikabirari001@gmail.com 2 Department of Computer Engineering, Late G. N. Sapkal COE, Nashik, jv.shinde@rediffmail.com Abstract- A sketch is freehand picture which serves various purposes such as tracing something that the artist can visualize so that they quickly explore concepts. Sketch-based communication is no newer, it is the oldest form of writing, and therefore Sketch-Based Image Retrieval (SBIR) can be a very valuable information search tool. Although sketch is good way to express people’s ideas, but sketches and photorealistic images are very different in the term of appearance. The photorealistic image shows attention to realistic detail. There is a large appearance gap in user sketches and photorealistic images, when people sketch, they usually focus on main structure or shape of object and only draw the semantic contour boundary. On the other hand the color, texture and shape of an object are main elements of photorealistic images, which makes it very difficult to directly match a sketch and the corresponding photo-realistic image. Therefore, to bridge this gap is fundamental challenge in SBIR. The existence of noisy edges on photo realistic image degrades retrieval performance and to bridge this gap there is framework consisting of line segment descriptor and noise impact reduction algorithm. Proposed descriptor extracts edges and captures the relationship between the edges. Object boundary selection algorithm used to reduce the noisy edges for which the hypothesis is used to maximize retrieval score, for which multiple hypotheses are generated. Keywords-descriptor, sketch retrieval, edge based, histogram, line relationship. I. INTRODUCTION A style of painting that contains an attention to realistic detail. Although edge extraction can bridge the appearance gap between sketches and photo-realistic images to some extent, it is quite common for noisy edges from background clutter, object detail and texture to be extracted with the object shaping edges. These noisy edges usually widen the appearance gap and degrade retrieval performance. Therefore, retrieval performance can be enhanced if the impact of noisy edges is reduced. Retrieval performance of the human visual system is not sensitive to these noisy edges since humans are able to distinguish object boundaries or contours from noisy edges based on their inference ability. Using this fact, algorithm can select the object boundaries from all extracted edges, the appearance gap can be filled and the performance of SBIR can be improved. This motivation provides with a new pathway to improved performance, which imposes a new requirement, i.e., that sketches/extracted edges should be treated as a set of lines, and the descriptors should be able to capture line-level features. This is because line-based descriptors give the flexibility to achieve edge selection or removal by setting the corresponding parts of the feature vector to a certain value, which is critical for boundary selection. Beside the need to solve the noise problem, that an effective descriptor for SBIR should be designed to describe lines and their relationships, rather than describing image patches, since a sketch/object boundary is essentially composed of lines (strokes) and the shape is determined by the relationships between these lines.