Multi-label Image Classification Based On Convolutional Neural Network

With the rapid development of computer vision and machine learning technology,and the continuous popularization of the Internet and its applications,a big data era of information explosion has come.Image as one of the most direct transmission medium,with the increasing demand for image applications,various types of image research topic come in a throng.Due to the wide application of image classification in various fields,such as military affairs,agriculture,public facilities and daily life,so how to classify the vast amount of image data efficiently and accurately is becoming especially important.Previous image classification is based on single-label image,that is,each image contains only one main foreground object,associated with only one label.However,in real life,an image often contains many different kinds of objects,so the classification of multi-label image is more practical.However,due to the problem of complicated layout and object occlusion,the task of multi-label image classification becomes more difficult.The application of artificial neural network in single-label image classification has been widely researched.Among them,the convolutional neural network has shown extremely excellent results and powerful performance.However,due to the complicated layout and the influence of light significantly improve the difficulty of classification,so in this case if the traditional convolution neural network framework is used to deal with multi-label classification tasks directly,it is obviously not able to solve the problem fundamentally.Therefore,how to apply convolution neural network to multi-label image classification task is still a problem to be explored.At first,through analyzing the relevant theories of object detection,multi-BING algorithm is proposed based on BING algorithm.Meanwhile,on the basis of multi-label classification theory,the multi-label image classification based on Fast R-CNN and Faster R-CNN is also comprehensively studied.The research results are as follows:(1)Object detection: to solve the shortcoming that BING algorithm models all objects as a model,multi-BING algorithm is proposed.First of all,the algorithm conducts K-means clustering algorithm on the training data by extracting the CS-LBP features.Then,BING feature model is established respectively for each kind of data.At the detection stage,the results are fused through a fusion strategy.The experimental results show that multi-BING algorithm has better detection effect than BING algorithm and OBN algorithm.(2)Multi-label image classification: through dividing multi-label tasks into several subtasks of single-label,on the basis of Fast R-CNN model,the object windows which obtained from the proposed multi-BING model are used as the input of Fast R-CNN model.On the basis of Faster R-CNN,according to the shortcomings that NMS algorithm forces the scores of the adjacent object windows to be zero,the proposed softer NMS algorithm is applied to Faster R-CNN model,which simultaneously retains the true windows and filters a great quantity of overlapping windows.In addition,different from traditional network model using Re LU function as the nonlinear activation unit,LRe LU function which can increase the number of neurons and modify the data distribution is applied to Fast R-CNN model and Faster R-CNN model.Thus,the classification results of the model are improved without increasing additional time and calculation complexity.The algorithm is designed by using the standard dataset PASCAL VOC2007.Firstly,we compare the performance of BING algorithm,multi-BING algorithm and OBN algorithm in object detection,DR is used as the evaluation criterion.Experimental results show that multi-BING algorithm can improve DR effectively under various overlaps and has an ideal detection effect.Then,we classify the multi-label image on Fast R-CNN model and Faster R-CNN model using the proposed algorithm,AP and m AP are used as the evaluation criterion.Experimental results show that the proposed algorithm can improve m AP effectively and reach an excellent classification effect.