Object Detection In Complex Background Based On Saliency And Template Matching

Object detection in complex background is one of the core issues in the field of computer vision.The main is the identification and localization of the objects in image.This paper proposed an object detection algorithm for complex background,which combines the saliency and template matching.Firstly,the image is pre-processed based on saliency.Secondly,the objects are detected in saliency region.Finally,the proposed object detection algorithm based on saliency and template matching is also experimentally analyzed.In the pre-process on the saliency,we firstly extract the eigenvector of the superpixel.And then the initial saliency trimap containing the foreground,background and unknown regions is obtained by random forest classification.The saliency detection is completed in the trimap.Then,the global saliency detection is realized by the highdimensional colour feature concept,combined by multiple traditional low-dimensional colour features,proposed in this paper.Meanwhile,the local saliency detection algorithm considering the features of adjacent superpixels is combined with the global detection.At last,the combination of these two saliency detections are iterated to obtain the optimal results of the saliency detection.The proposed object detection algorithm is based on the shape template matching in the salient region.We introduce a new shape descriptor to represent the contour shape for matching between the template image and the image to be tested.Moreover,the hypothesis of object detection is obtained by applying the dynamic time warping matching algorithm and depth-first search strategy.Finally,the hypothesis verification is performed to achieve the object detection in a complex background.The experimental analysis of the algorithm includes three parts: experiments of saliency region detection,final experiments of object detection,and performance analysis of saliency pre-process and shape descriptors.The saliency experiments are tested in MSRA-B,ECSSD and PASCAL-S datasets.We find that the detection accuracy in the adaptive threshold algorithm can reach 89.5%,which is significantly higher than the common GC and RBD models.And the time is reduced by 79.5%compared with the DRFI model with the guaranteed detection accuracy.The final object detection experiments are carried out in ETHZ,INRIA Horse and Caltech 101 datasets.When the false positive per image is 0.3,the average detection rate of the algorithm can reach 98.4%.And the average detection rate of the algorithm can reach 99.7% while the false positive per image is 0.4.The proposed algorithm performs better than other shape-based object algorithms.Finally,in the ETHZ dataset,it is proved that the saliency pre-processing are necessary and effective.At the same time,the shape descriptors are experimentally analyzed in two shape datasets: MPEG-7 and Kimia's.Compared with the common shape feature algorithm,the shape descriptors of this paper have the best performance in matching scores and retrieval numbers in image retrieval.