Video Shot Boundary Detection Algorithm Based On EMD And KNN

Nowadays our world is coming into the age of information, everyday we will get lots of various data which are full of information, most of the data are video and image. Scientific research tell us people get 70% information from eyes, so the research of video and image is always hot. The amount of information which video carry is far more than audio and other data, the video information has lots of good features such as radication intuition materiality vividness and great efficient, this made the video transmission the most leading ways to transmit information. Along with the enrichment of video data, we need to find the part of video which we interested in, this requirement has been more and more exigent. People hope in the very future we can use the video data easily and quickly just like we use document data now, so we need to establish video database and through some meaningful ways to index and to help us to arrange and organize the database. Content based retrieval is just the key technology in the video database.Because video data is always rich in content, it is difficult to directly used for indexing, so we need video data segmentation. Video boundary detection technology is used to segment video section which contains a wealth of information, the process is to find out the shear and graded, on the base of that we can segment video section which has some different scene. Video data is various, the lens of video data has lots of effects, A lot of factors are between the front and rear frames, this is enough to cause some interference, such as abrupt changes in light and movement, and other large objects. These are the sensitive factors of border detection algorithm, making the algorithm easy for mistakenly seized and missed, so the issue for minimizing interference and enhancing the robusticity of edge detection algorithm is very important.The so-called content-based video retrieval is a process that according the Color Texture and shape characteristics of designated targets data to identify the corresponding video or image data. Effective and rational segmentation is the base work of this process. Video data is posed by frames, shot is posed by a string of consecutive frames, it is the basic structure of video data unit, normally used for presenting a continuous movement in Time and space. Scene is posed by shot which is logically related, it describes a complete event. Video Database Application is not a simple keyword search, which is popular in traditional database, the content-based video Retrieval can search the database according to the content of specific targets. Content-based video Retrieval can find corresponding image or video clips from the database which have formulation characteristics of the specific targets. The general work of video segmentation includes key frame extraction, video ACRONYM, and video identification. As the video is a hierarchical structure model of scene, camera, and frame, lens is a series of associated frame, so lens is the most basic units of video production, editing, indexing, and video boundary detecting is the most important work of video segmentation. The so-called video shot boundary detection is to identify the initial and end frame in a video camera. The video data is diversity and richness, and sometimes for the goal of achieving certain effects, the conversion of type is very different between frames, such as shear, which means there is no time delay between lens, and gradual transition changes gradually, There are many types of gradual change. At present, the video lens shear detection has got considerable positive research result. Gradual change detecting is more complex, so it is hard to get the same research result as shear detectingVideo shot boundary detection need comparative analysis of continuous frame. There are three main methods, named pixel-based method, block-based method, and histogram-based method. The pixel-based method need to calculate the difference of the corresponding pixels, and the number of pixels is very large, so the calculation will be time consuming, at the same time this method just overlook local movement in the shot, so it is very sensitive to the local movement. The histogram-based method takes the histogram of each frame as feature, it describe the gray structure of whole image, so it is not limited to single pixel and that made the method not sensitive to local movement. In another word this method will make histogram-based method a robust method and make little mistake, but this method is base on the whole image, it doesn't tell the distribution of gray in extensity, if there are two images which are similar in gray histogram but different in extensity, then that method will mistaken the two images as in the same shot, this maybe overlook shear. On the base of histogram-based method, there is a method called block-based method, this is proposed to solve the disadvantage of histogram-based method, this method block the video image in extensity first, then use histogram-based method to calculate every block. This process consider the video image a whole image and at the same time consider the gray histogram different in extensity, it is good for boundary detection.In this paper, in the big background of information age, it introduced a content-based retrieval system, and the concept of video data, including video data in compression domain, then leads to the video shot boundary detection algorithm. It presents several important concepts of video data. And discuss the significance and some problems of detection methods, because the algorithms of video boundary detection based on the pixel block and histogram methods are all sensitive to light, the paper introduced the EMD algorithm idea. Through the research of EMD decomposition method, we found that the lowest frequency component of IMF represent the trend and evenness of original signal, as to video image, the minimum frequency IMF reflects the light and energy distribution of the image, so we use DEMD to decompose frame, after the process of getting the IMF component, just replace the old frame sequence with new frame sequence which is wiped off light energy, on the base of that use histogram-based method to calculate the interference. There are still some interference, it need filtering process to avoid the interference, and make the signal clear and simple, it also need recognize the signal and make sure which transformation has happened.There are various methods of feature extracting that using large number of video data, the result shows that using the U component of the model YUV as the feature to computing the difference between frames is more efficient, it can avoid the sensitive to the object movement while lots of shot boundary detecting algorithms always meet, and it is more efficient to detect the gradual transition, on the base of that, we use flexible window to filter the histogram, then make the graph in the histogram to clear signal features, after that we can make use of KNN sorting algorithm to classify the character graph. KNN classification algorithm is concept of machine learning algorithm, through training it can constantly increase its knowledge reserves, this algorithm calculate the difference between new object and old examples, after that it will judge which category the new object belong to, according research, the shear curves have similar characteristics to the rectangular-shaped curve; and graded curves have similar characteristics to the triangle-shaped curve. Therefore, if we do proper classification with curved graph after filtering, and identify each type of graphics effectively, we can realize the shear and gradual. KNN algorithm can recognize different signals after training, and classify the signal into some class, then detect the shot boundary of video data. This method dispose all sample in a computer, decision-making base on all the sample, it need calculate all the distances between new object x and every old samples N, so this method need great amount of computation.