基于空间稀疏编码的MIL算法及刑侦图像分类

Spatial Sparse Coding Based MIL Algorithm for Criminal Investigation Image Classification

针对刑侦图像分类问题,提出一种基于空间稀疏编码(SSC)的多示例学习(MIL)算法。首先,利用稠密尺度不变特征转换(SIFT)原理设计一种带有示例位置信息的多示例建模方案,将刑侦图像分类问题转化为MIL问题;然后,基于多样性密度(DD)函数及稀疏编码(SC)理论,设计了一种针对MIL的字典构造方法及空间稀疏编码方案,用于计算多示例包的元数据(metadata);最后,结合大尺度线性支持向量机方法,提出了一种SSC-MIL的MIL新算法。14类真实刑侦图像的对比实验表明,该算法是有效的,且分类精度高于其他方法。

Focusing on the classification problem of the criminal investigation,a multi-instance learning(MIL)algorithm based on spatial sparse coding(SSC)is proposed.By using the dense scale invariant feature transform(SIFT)principle,a multi-instance modeling scheme with instance position information is constructed to transform the problem of criminal investigation image classification into a multi-instance learning(MIL)problem.Based on the diversity density(DD)function and the sparse coding theory,a new dictionary construct method and spatially sparse coding(SSC)technique are designed for MIL,to extract the metadata for each multi-instance bag.At last,a new MIL algorithm called SSC-MIL is proposed by combining the large-scale linear support vector machine method.Experimental results on the 14 cases of real criminal investigation image show that the proposed method is effective,and the classification accuracy is higher than other methods.