Research And Application Of Automatic Image Annotation Algorithm

As the quantity of image data presenting geometric growth, the existing image retrievalproblems which are caused by its using keywords or text label as retrieval mechanism arestrongly appeared. Besides, artificial job on the annotation image is laborious and time-consuming, and artificial annotation has subjective and uncertainty nature which lead to onepossibility: the annotating results made by different individuals can exist kind of differenceamong these results. Hence, high efficiency and precise method for automatic image annotationsurelycan efficientlypromote the performance of image-annotation system.For previous one kind of most normal models in automatic image-annotation field，generative model, this article combining with the method would study the potential factors thatcause slower speed in the process of automatic annotation. And as we know, traditionalgenerative model can divide into three aspects：text representation, image representation andestablishment of image-text theme. Moreover, all of these above means its showing one piece ofpicture relative to text matrix and the bottom features and then it correlates these two specificrepresentations with each other in something method to produce one theme which can be usedfor classification annotation of subsequent images.For some disadvantages existing in traditional local feature extraction algorithm, such asthe inefficiency when extracting feature points and the cost when calculating main direction ingenerating descriptor, etc., an improved accelerated radial SURF algorithm is proposed in thispaper, which speeds up the process of local invariant feature extraction in the circumstance ofless precision loss as much as possible with combination of both merits of SURF algorithm andRGT (Radial Gradient Transform). In the detection step, exactly the process of feature pointlocation, calculating response layers in dimension space construction are decreased, andcalculation course of corresponding point is placed in detection phase. In the feature descriptionstep, the process of determining feature orientation is canceled, and RGT transformation isconducted to Haar wavelet response in surrounding regions of feature point, then these regions are divided into concentric circles and response results are synthesized, finally rotation-invariantfeatured descriptors are achieved with those responding results. AR-SURF algorithm is provedthat it can decrease the loss of time and space, increase the speed of detection and improve theextracting effect, so it is more suitable mass image processing than SURF.When the article indicates the process of making Image-text correlative theme，thereselects LSA algorithm and CCA algorithm which are good at precision and speed, meanwhile,uses the method of sparse vector to promote the CCA’s speed in order to get SparseCCAalgorithm, and then we can make the new frame of automatic image annotation. In theory, theannotation based on the automatic image annotation model can work much better results out atthe aspects of image mining and image retrieval.