Condensed semantic tree model for image category representation

Abstract

This paper presents a condensed semantic tree model for representing image category. For a specific application area, a semantic concept space is defined. According to the annotation for an image, a real-value semantic vector is gained that describes the content of it. In order to represent image category, condensed semantic tree model is introduced. It is a triple level structure. The bottom level is a semantic concept mask, which selects those concepts relevant to semantic category. The middle level is composed of three semantic modules, which extract high-level semantic of an image. The top level analyzes the probability that an image is belong to a specific image category. Every semantic category has different model configuration. The experimental results illustrate that the effectiveness of the proposed condensed semantic tree model is good.

Human Behavior Recognition based on Conditional Random Field and Bag-Of-Visual-Words Semantic Model

“…We use HCRF tool bag, features extracted by VC ++and Matlab completed the training and testing. Video data come from the Weizmann Human Action Dataset [15] and network video. Our algorithm has semantic mark, training is divided into three parts, the upper layer complete the primary semantic mark model training, middle layer complete the middle semantic mark training.…”

“…1) semantic dictionary constructed by Random Forest algorithm [7] which extract visual primitives to classify the underlying visual vocabulary and vocabulary again intermediate classify, integrate context and spatial feature [13][14][15], then we will build a visual semantic dictionary with high degree distinction;…”

“…In Figure 5, the semantic model comprising: body parts model and the human body behavior model. Due to some words of gymnastics semantic vocabulary has high appearing probability, therefore, in this paper on the basis of [15] they weighted distribution the semantic vocabulary we proposed to removal of the scene semantic model and only weight distribution the human body parts visual vocabulary and human behavior related visual vocabulary. Any image C in data images set V, if the visual vocabulary probability of the body parts is P 0 human behavior related visual vocabulary probability is P , by the formula P C = P 0 + P , α + β = 1,we can calculate the error rate of C .…”