Research Of Large-scale Image Annotation Methods

Automatic image annotation is the key technology to retrieve、manage and browsethe image data. In the reality of the Internet environment, the research of automatic imageannotation is more concerned that in the increasing massive image data case, how to usea variety of machine learning, pattern recognition and statistical modeling to realizeautomatical image annotation of large-scale data have great significance.To further improve the accuracy and efficiency of labeling, considering theinfluence of calculating large data sets, and the complexity of large-scale imagescharacteristics, This paper has presented classification learning method based-on randomforest and large-scale training dataset annotation method based on a distributed labelpropagation. classification learning method based-on random forest is consideredpreliminary image classification, the sample tree leaf nodes in the same set of propertieshave the same locality, but because the picture is a multi-labeled training set, so proposedthe establishment of a random pre-classification tree. For large-scale data, the paperproposed proposed a method of integrated random tree. Choose a smaller variance leafnode as an input distributed computing proposed sequencing map task distributed method.large-scale training dataset annotation method based on a distributed label propagation isa parallel computing method, distributed knn proposed based label propagation method,each map task corresponds to a leaf node set, while optimizing processing temporary dataand computing results at every map task.In order to verify the feasibility and effectiveness of the proposed algorithm, makeexperiments on distributed clusters. Experimental results show that the improved randomforests algorithm can classify the training samples, meanwhile its pretreatment imagesort, can well reduce the amount of computation behind. Use mapreduce method canimprove the efficiency of labeling the new picture. At the effective time, the picture hasimproved precision and recalling rates.