Image Retrieval Based On Shape Features Under Complex Conditions

Image retrieval has always been a research hotspot in the field of multimedia technology,and it's also widely applied in search engines,trademark search and e-commerce trade,etc.The application of this paper is the bottle retrieval under the e-commerce trade.Since the bottle can be colored later,the shape becomes the most important distinguishing feature,while segmentation is a necessary step for the shape feature extraction.The data sets targeted here include database image and retrieve image.The main challenges of the segmentation of database image are gradient background noise,horizontal line noise,and weak edge.Due to the smoothness,transparency of the bottle and uneven illumination of environment,etc.,the segmentation of retrieve image often have very complex environment: weak edges,missing edges,extreme uneven grayscale caused by transparency and shadows of bottle,and highlights produced by reflection and refraction.While the shape retrieval faces the problems of large intra-class distance but small inner-class distance.Besides,for a retrieve image,our actual application system requires returning the first 16 similar pictures within 30 seconds.Our work is divided into three steps: database image segmentation,retrieve image segmentation and shape retrieval,each of which to solve the above problems respectively.The main contributions of this paper are as follows:(1)The gradient background noise and weak edge problems of database image segmentation are solved by improving the region growing method.Foreground probability map is used to adaptively scale the criterion of region growth,which not only preserves the characteristics of the region growing method for filtering the gradient background noise,but also improves the over-segmentation problem.The closed operation is used to process the horizontal line noise,and finally the segmentation of the database image is completed.(2)For the problem of highlight in retrieve image,we propose a general highlight detection method for both reflection and refraction.This method initializes and scales the highlight area according to the similarity of brightness and region structure between highlight of reflection and refraction,and then the brightness of the detection area is changed to remove the highlight.The algorithm serves as a pre-processing step for retrieve image segmentation to alleviate the problem of gray unevenness.(3)For the time requirement and complex segmentation conditions of the retrieve image,a one-step optimization GAC algorithm based on graph cut is proposed.This algorithm discretizes and maps the GAC model based on GVF field into the boundary term in the graph,and uses the improved region growing method to generate the initial probability map as the initialization of the region term in the graph,which improves the severe gray scale unevenness and makes the whole model can be optimized in one step.In the retrieve image dataset,this algorithm reached 91.03% MIo U,which is 36.36% higher than One Cuts and 11.23% higher than GAC model,and the segmentation time of our algorithm is 4.5 times faster than One Cuts and at least 384 times faster than GAC model.(4)For the difficulties and time limitation of this retrieval problem,we propose a Bo W retrieval model based on the improved SC feature.Firstly,the SC feature based on normalized residual coding is proposed,the shape is more accurately represented by the normalized two-dimensional residual of the sample point relative to the center of the grid.Secondly,the contour is divided into multiple contour fragments,for the all contour fragments of training images,the above improved SC are extracted to perform clustering,coding,pooling and SVM training.Thirdly,the supervised weights of SVM are used to generate weights for the pooled features,and finally the pooled features are weighted to obtain shape features that can be used for quick image retrieval.This method achieves 65.62% m AP on the retrieve image dataset,which is 42.64% higher than the second-ranked retrieval method.This work provides a feasible solution for solving the core technical problems of complex image segmentation and rapid shape retrieval with small intra-class distance,which lays a foundation for the design of a practical bottle retrieval system.