Video Caption Detection Algorithm Based On Multiple Instance Learning

In recent years, with the rapid development of Internet and multimedia, more and moreimages or video information appear, but there is an insurmountable gap between the bottomfeatures of the image and high-level semantics, video Caption can effectively takehigh-level semantic of images and gets more useful information. Most researchers mainlyextract subtitles by low-level features of characters, such as, shape, texture and so on, but itsaccuracy and efficiency should be improved. Machine learning plays an important onextracting the subtitles in image processing.This thesis proposes a video caption extraction algorithm based on multi-instancelearning. Multi-instance learning is a relatively new machine learning framework, changes thetraditional classification model. The data set is divided into packages with marked andwithout marked, if the package has at least a positive example, then mark the bag packagepositive, the contrary, if all of the instances in the bag are negative, mark it negative, and thencontinue to re-built packages to examples in positive bags, iterate, until separate positive andnegative examples at the greatest degree. This thesis proposes a new method through animproved multi-instance learning algorithm SVM-MI, because support vector machine is builton the theory of VC dimension and structural risk minimization principle. Based on thelimited sample information, seek the best compromise between model complexity andlearning ability, in order to obtain the best generalization ability. SVM-MI algorithm solvesthe high dimensional and nonlinear problems by Support Vector Machine; succeed to convertmulti-instance learning into learning, it has significantly improved in time complexity androbustness than the existing multi-instance learning algorithm. The algorithm also achievesbetter results in the video caption extraction applications. Experiments show that accuracy ofcaption extraction has been greatly improved.