Research On Apple Lesion Retrieval Based On Video Shots

The thesis did some preliminary research on key technologies common used in content-based video retrieval including video shot boundary detection, feature extraction, shot clustering and similarity measurement which will be applied to retrieval apple lesion images in video shots of apple extermination of disease and insect pest, providing a new means to spread the academic achievements of apple extermination of disease and insect pest. To summarize, there are mainly four contributions:(1) The typical algorithms of video shot boundary detection, including the algorithm based on pixel comparison, the algorithm based on block comparison, the algorithm based on histogram comparison and the algorithm based on DC coefficient, are compared through experiments in order to get the best one that will be applied to the targeted videos and the last algorithm is selected due to best performance as a result. In cut transition detection, the average precision and recall are 96.03% and 91.95% and in gradual transition detection, that is 74.19% and 80.62% respectively.(2) Feature extraction and clustering of video shots of apple extermination of disease and insect pest. In this paper, gray histograms and color histograms (quantized to 512 bins) of video frames are computed respectively firstly, which are used to calculate the average gray histograms and color histograms of video shots as their color characteristics. Then the video shots are clustered using Kmeans based on the color characteristics. As a result, the shots are clustered into 6 and 7 classes based on gray and color histogram respectively. The experiment finds that video shots are clustered efficiently using Kmeans.(3) Similarity measurement. For gray histogram and color histogram, the retrieval performances of Manhattan distance and Euclidean distance are compared respectively. The experimental result shows that using color histogram as the retrieval feature, Manhattan distance for the right shot class and Euclidean distance for the similar shots performs best. When searching the right shot class, the right rate can reach as high as 92.44%. When comes to the similar shots, if only one shot is got in the result, the right rate is 80.59%. While if the shot number of the result grows to 5, the right rate goes up to 89.84%.(4) EmguCV and DirectShow .NET toolkits are employed based on Visual Studio 2008 to develop an apple lesion retrieval system based on the video shots using C#.The experimental results have demonstrated that all these methods are satisfying. Video shots with apple lesion images can be retrieved efficiently, thus popularizing the research results of apple extermination of disease and insect pest.