基于深度特征选择的多示例算法及在图像分类中的应用

Multi-instance learning based on deep feature selection and its application to image classification

信息技术高速发展为人们生活带来便利的同时,海量的信息也给人们带来许多困扰,如图像检索变得越来越困难。因此智能化地进行图像分类识别具有重要的研究意义。基于多示例学习的图像分类方法得到了越来越多学者关注,。也提出了一些算法,但仍存在特征表达有限,模型受无关示例影响较大的问题。文章提出一种基于深度特征选择的多示例算法,并验证了此算法的有效性。该方法首先利用深度预训练模型提取示例高层语义特征,再将包投影到示例获取图像的深度特征,然后通过特征选择剔除干扰示例的影响,最后利用训练好的SVM分类器对图片类别进行预测。不同数据集上的实验结果表明,该方法有效地实现了图像分类。

The rapid development of information technology brings convenience to people’s lives.At the same time,massive information also brings many troubles to people,such as inefficient image retrieval.Therefore,intelligent image recognition and retrieval becomes important.Image classification methods based on multi-instance learning(MIL)have drawn increasing attention from scholars.Some effective algorithms based on MIL have been put forward,but the image representation is weak and the model cannot well deal with the disturbance of the useless instance.Therefore,this paper proposes a multi-instance algorithm based on deep bag feature selection and verifies the effectiveness of the algorithm.In this method,the pre-training deep network is firstly used to extract the instance high-level features and then the deep bag representations are computed by mapping to all instances.To elimination the influence of the noisy instances,an efficient feature selection method based on l2,1 is introduce to get the more discriminative bag mapping features.Finally,a learned SVM is used to predict the image label.Experimental results on different data sets show the effective of the proposed method.