基于条件随机场的多标签图像分类识别方法

Multi Label Image Classification and Recognition Method Based on Conditional Random Field

针对图像分类识别效率低、精度不高等问题,提出一种基于条件随机场的多标签图像分类识别方法。转换图像特征语义化,将特征连接距离与信息数量相结合,获得像素关联概念语义相似度,运用互信息评估方法表明语义转换的有效性,提高图像分类精度;建立条件随机场模型,得到图像特征语义序列与图像标签间的映射关联,有效挖掘标签间的语义相关性;利用欧氏距离动态损失函数,增强图像识别预测值准确性,将二维样本特征变换成包模式,得到簇图像内多个目标的差异性,通过数据集分类学习实现多标签图像的精准分类识别。仿真结果证明,提出的方法对标签图像的分类识别效果较强,且耗时较短。

Due to low efficiency and low precision of image classification and recognition,this paper puts forward a method of multi-label image classification and recognition based on conditional random field.Firstly,we transformed the image feature semantically,and combined feature connection distance with information quantity to obtain the semantic similarity of pixel association concept.Secondly,we used the mutual information evaluation method to prove the effectiveness of semantic transformation,and thus to improve the accuracy of image classification.Thirdly,we built the conditional random field model and thus to get the mapping association between image feature semantic sequence and image tag,so that we could effectively mine the semantic correlation between tags.Moreover,we used Euclidean distance dynamic loss function used to enhance the accuracy of prediction value of image recognition.Then,we transformed two-dimensional sample features into packet mode,and thus to get the differences among targets in cluster image.Finally,we completed accurate classification and recognition of multi-label images through data set classification learning.Simulation results show that the proposed method is effective for label image classification and recognition.