公开(公告)号：US09137529B1

公开(公告)日：2015-09-15

申请号：US14208352

申请日：2014-03-13

Applicant: Google Inc.

Inventor： Michele Covell ,  Mei Han ,  Saurabh Mathur ,  Shumeet Baluja ,  Vivek Kwatra

IPC: G06F15/18 ,  H04N19/105 ,  H04N19/17 ,  H04N19/189

CPC classification number: G06K9/6202 ,  G06F17/30247 ,  G06K9/00684 ,  G06K9/4604 ,  G06K9/6267 ,  G06K9/6292 ,  G06K9/66 ,  G06T1/0021 ,  H04N19/105 ,  H04N19/17 ,  H04N19/189

Abstract: An exemplar dictionary is built from exemplars of digital content for determining predictor blocks for encoding and decoding digital content. The exemplar dictionary organizes the exemplars as clusters of similar exemplars. Each cluster is mapped to a label. Machine learning techniques are used to generate a prediction model for predicting a label for an exemplar. The prediction model can be a hashing function that generates a hash key corresponding to the label for an exemplar. The prediction model learns from a training set based on the mapping from clusters to labels. A new mapping is obtained that improves a measure of association between clusters and labels. The new mapping is used to generate a new prediction model. This process is repeated in order to iteratively refine the machine learning modes generated.

Abstract translation: 由数字内容的示例构建示范字典，用于确定用于对数字内容进行编码和解码的预测器块。 示范字典将样本组织成类似样本的集群。 每个集群映射到一个标签。 机器学习技术用于生成用于预测样本的标签的预测模型。 预测模型可以是哈希函数，其生成与样本的标签相对应的散列密钥。 基于从集群到标签的映射，预测模型从训练集学习。 获得了一种新的映射，改进了集群和标签之间的关联度量。 新映射用于生成新的预测模型。 重复该过程以便迭代地改进所生成的机器学习模式。

公开(公告)号：US08855437B1

公开(公告)日：2014-10-07

申请号：US13907000

申请日：2013-05-31

Applicant: Google, Inc.

Inventor： Jing-yu Cui ,  Vivek Kwatra ,  Michele Covell ,  Mei Han ,  Saurabh Mathur

IPC: G06K9/46 ,  G06T9/00

CPC classification number: G06T9/004 ,  G06K9/6219

Abstract: Compression of an image is performed based on prediction of target blocks of an image from candidate source blocks of the image. Heuristics are used for identifying the candidate source blocks, for example, source blocks are selected from within a cluster of similar blocks obtained by K-means clustering. For each target block, a region adjacent to the target block is identified and a set of candidate source blocks along with candidate source regions adjacent to the candidate source blocks are identified. The candidate source regions are ranked based on the differences between the candidate source regions and the target source region. Each candidate source block is described using its rank and residual information describing differences between the candidate source block and the target block. The candidate source block that can be described using a minimum amount of information is selected for predicting the target block.

Abstract translation: 基于来自图像的候选源块的图像的目标块的预测来执行图像的压缩。 启发式用于识别候选源块，例如，从通过K均值聚类获得的相似块的簇内选择源块。 对于每个目标块，识别与目标块相邻的区域，并且识别一组候选源块以及与候选源块相邻的候选源区。 基于候选源区域和目标源区域之间的差异对候选源区域进行排序。 使用其等级和残差信息描述候选源块和目标块之间的差异来描述每个候选源块。 选择可以使用最小量的信息描述的候选源块用于预测目标块。

公开(公告)号：US08787692B1

公开(公告)日：2014-07-22

申请号：US13946965

申请日：2013-07-19

Applicant: Google Inc.

Inventor： Mei Han ,  Michele Covell ,  Saurabh Mathur ,  Shumeet Baluja ,  Vivek Kwatra

IPC: G06K9/36

CPC classification number: G06K9/6219 ,  G06K9/6807 ,  H04N19/30 ,  H04N19/90

Abstract: An exemplar dictionary is built from example image blocks for determining predictor blocks for encoding and decoding images. The exemplar dictionary comprises a hierarchical organization of example image blocks. The hierarchical organization of image blocks is obtained by clustering a set of example image blocks, for example, based on k-means clustering. Performance of clustering is improved by transforming feature vectors representing the image blocks to fewer dimensions. Principal component analysis is used for determining feature vectors with fewer dimensions. The clustering performed at higher levels of the hierarchy uses fewer dimensions of feature vectors compared to lower levels of hierarchy. Performance of clustering is improved by processing only a sample of the image blocks of a cluster. The clustering performed at higher levels of the hierarchy uses lower sampling rates as compared to lower levels of hierarchy.

Abstract translation: 从用于确定用于对图像进行编码和解码的预测器块的示例图像块构建示范字典。 示例性字典包括示例图像块的分级组织。 通过例如基于k均值聚类来聚类一组示例图像块来获得图像块的分级组织。 通过将表示图像块的特征向量变换为较少的维度来提高聚类的性能。 主成分分析用于确定尺寸较小的特征向量。 在层次较高的层次上执行的聚类与较低级别的层次相比，使用较少的特征向量维度。 通过仅处理集群的图像块的样本来提高聚类的性能。 与较低级别的层次相比，在较高层次上执行的聚类使用较低的采样率。