SBIR Phase I: Automated Image Annotation

这个小型企业创新研究（SBIR）第一阶段项目通过使用基于实例的学习方法、内容分析和预注释的图像来检验自动为新图像分配关键词的可行性。该项目将提供一种新的颜色量化方法，使用一组常见的代表性和歧视性颜色来描述图像。这减少了基于实例的方法在确定要分配哪些关键字时所使用的距离度量所需的计算次数。该项目还将研究将低级颜色域转换为新的特征域的效用，有效地去除颜色特征之间的相关性。当基于实例的方法使用不考虑相关性的距离度量时，这减少了注释错误的可能性。用户在使用相关性反馈执行搜索时隐式地改进这些初始注释。通过检查用户反馈，分配给图像的关键字的相对重要性被修改。相关程度越高的关键词权重越大，相关性越低的权重越小。这允许忽略错误分配的关键字。因此，该系统为自动为图像分配关键字提供了一种有效的方法，并允许通过合并用户搜索结果进行自动更正。该项目将为参与图像生成和/或收集的组织提供方便和廉价的注释新图像。在货币和效率方面都实现了节约。通过自动化注释过程，对专门检查和分类原始数据的人员的需求大大减少，从而降低了成本。此外，通过自动化流程，可以提高分配注释的速度，从而实现更大的吞吐量。此外，通过注释图像，基于关键字的搜索和检索系统现在可以在组织的图像集合上使用，允许对现有软件产品进行更大的利用，并允许对集合进行更大的曝光和利用。

This Small Business Innovation Research (SBIR) Phase I project examines the feasibility of automatically assigning keywords to new images using instance-based learning methods, content analysis and pre-annotated images. This project will provide a new color quantization method that describes images using a common set of representative and discriminatory colors. This results in a reduction of the number of calculations required by the distance metrics employed by instance-based methods in determining which keywords to assign. This project will also investigate the utility of transforming the low-level color domain into a new feature domain, effectively removing the correlations between color features. This reduces the potential for annotation error when the instance-based methods employ distance metrics that does not account for correlations. Users implicitly refine these initial annotations when they perform searches using relevance feedback. By examining the user feedback, the relative importance of the keywords assigned to an image is modified. The more relevant keywords are assigned greater weights and the less relevant smaller weights. This permits erroneously assigned keywords to be effectively ignored. Hence, this system provides an efficient means for automatically assigning keywords to images and allows for automatic corrections by incorporating the results of user searches.This project will provide organizations that are involved in image generation and/or collection to easily and inexpensively annotate new images. Savings are achieved in both the monetary and efficiency arenas. By automating the annotation process, the need for a staff dedicated to examining and categorizing raw data is greatly reduced, resulting in reduction of costs. Furthermore, by automating the process, the speed at which annotations can be assigned is enhanced, allowing for greater throughput. In addition, by annotating the images, keyword-based search and retrieval systems can now be used on the organization's image collection, allowing for greater leverage of existing software products and permitting greater exposure and utilization of the collection.