時(shí)空視頻檢索

前言

The problem of semantic video scene eategorisation by using spatio-temporal information is one of the significant open challenges in the field ofvideo retrieval. During the past few years, advances in digital storagetechnology and computer performance have promoted video as a valuableinformation resource.Numerous video retrieval techniques have beensuccessfully developed. Most of the techniques for video indexing andretrieval have extended the previous work in the context image basedretrieval. In this process, video sequences are treated as collections of stillimages. Relevant key-frames are first extracted followed by their indexingusing existing image processing techniques based on low-level features. Forthe research in the book the key question is how to encode the spatial andtemporal information in video for its efficient retrieval. Novel algorithms areproposed for matching videos and are compared them with state-of-the-art.These algorithms take into account image objects and their spatialrelationships, and temporal information within a video which correlates withits semantic class. Also, the algorithms perform hierarchical matchingstarting with frame, and shot level before overall video level similarity can becomputed. The approach, then, is exhaustively tested on the basis ofprecision and recall measures on a large number of queries and use the areaunder the average precision recall curve to compare the methods with those inthe literature. As a part of this book an international video benchmarkMinerva was proposed on which the results have been discussed.

內(nèi)容概要

本書(shū)重點(diǎn)挖掘了視頻的時(shí)空關(guān)系，探索了利用機(jī)器學(xué)習(xí)的方法進(jìn)行視頻切割、語(yǔ)義分類(lèi)。本書(shū)分七章，闡明了圖像的各種特性，論述了視頻的特征，系統(tǒng)介紹了視頻的時(shí)空邏輯關(guān)系、視頻的統(tǒng)計(jì)分析方法，研究了如何捕捉視頻的時(shí)空特性，如何利用人工智能神經(jīng)網(wǎng)絡(luò)進(jìn)行視頻切割，如何訓(xùn)練計(jì)算機(jī)“學(xué)會(huì)”用人類(lèi)的思維進(jìn)行視頻語(yǔ)義分類(lèi)、檢索。各章節(jié)撰寫(xiě)排列體現(xiàn)了從簡(jiǎn)到繁、由淺入深、從理論到實(shí)際、從技術(shù)到系統(tǒng)的特點(diǎn)。    本書(shū)可以作為高等學(xué)校信號(hào)與圖像處理、計(jì)算機(jī)科學(xué)、機(jī)器學(xué)習(xí)、人工智能、機(jī)器視覺(jué)等領(lǐng)域的研究生教材和參考書(shū)，也可以作為在這些領(lǐng)域從事相關(guān)工作的高級(jí)科學(xué)技術(shù)人員的參考書(shū)。

書(shū)籍目錄

Chapter I  Introduction  1.1  Motivation  1.2  Proposed Solution  1.3  Structure of BookChapter lI  Approaches to Video Retrieval  2.1  Introduction  2.2  Video Structure and Properties  2.3  Query  2.4  Similarity Metrics  2.5  Performance Evaluation Metrics  2.6  SystemsChapter IlI  Spatio-temporai Image and Video Analysis  3.1  Spatio-temporal Information for Video Retrieval  3.2  Spatial Information Modelling in Multimedia Retrieval   3.3  Temporal Model   3.4  Spatio-temporal Information FusionChapter IV  Video Spatio-temporal Analysis and Retrieval (VSTAR). A New Model   4.1  VSTAR Model Components  4.2  Spatial Image Analysis  4.3  A Model for the Temporal Analysis of Image Sequences    4.4  Video Representation, Indexing, and Retrieval Using VSTAR  4.5  Conclusions  Chapter V  Two Comparison Baseline Models for Video Retrieval  5.1  Baseline Models    5.2  Adjeroh et al. (1999) Sequences Matching--Video Retrieval Model    5.3  Kim and Park (2002a) data set matchingmVideo Retrieval Model  Chapter VI  Spatio-temporal Video RetrievalmExperiments and Results    6.1  Purpose of Experiments    6.2  Data Description    6.3  Spatial and Temporal Feature Extraction   6.4  Video Retriewd Models: Procedure for Parameter Optimisation  6.5  Video Retrieval Models:Results on Parameter Optimisation  6.6  Comparison of Four Models  6.7  Model Robustness (Noise)  6.8  Computational Complexity  6.9  ConclusionsChapter VII  Conclusions  7.1  Reflections on the book as a whole……Reference

章節(jié)摘錄

插圖：In Ioka and Kurokawa （1992）, the user is allowed to specify a query by drawing amotion trajectory. The similarity is computed as the Euclidean distance between thequery vector and the stored vector for each given interval to match the specifiedtrajectory with the trajectories of the sequences in the database.3.3.2.2 Correlation Based ComparisonThis approach is based on finding the maximum correlation between tile predictorand the current one, for gesture recognition to identify actions. Martin and Shah （1992）used dense optical flow fields over a region, and computed correlation between differentsequences for matching.  In Campbell and Bobick' s  （1995）  work on gesturerecognition, the learning/training process is accomplished by fitting the unique curve ofa gesture into the subset of the phase space with low-order polynomials.Rui and Anandan （2000） addressed the problem of detecting action boundaries ina video sequence containing unfamiliar and arbitrary visual actions. Their approach wasbased on detecting temporal discontinuities of the spatial pattern of object region motionwhich correspond to the action temporal boundary to capture the action.  Theyrepresented frame-to-frame optical flow in terms of the coefficients calculated from all ofthe flow fields in a sequence, after principal components analysis to determine the mostsignificant such flow fields. The temporal trajectories of those coefficients of the flowfield are analysed to determine locations of the action segment boundaries of videoobjects.

編輯推薦

《時(shí)空視頻檢索(英文版)》：學(xué)者書(shū)屋系列

圖書(shū)封面

評(píng)論、評(píng)分、閱讀與下載

---

    時(shí)空視頻檢索 PDF格式下載

---