Research And Application Of Multi-Label Image Model With Confidence

With the rapid development of information and multimedia technologies,the image has shown an exponential growth,and thus how to deal with image classification has become a problem to be urgently solved.In real life,an image is often described by multiple labels simultaneously,and the image classification is a typical multi-label problem.Moreover,In high risk areas,a classification failure will result in serious consequences.In this paper,convolutional neural networks(CNN)are introduced into the framework of inductive conformal predictor(ICP)learning to propose a multi-label inductive confidence prediction model based on CNN(MLICP-CNN),which could make an calibrated confidence evaluation on the outputs.The MLICP-CNN model was finally applied to the prototype system of chest X-ray for multi-label diagnosis,which can output a confidence diagnosis.The major research contents of this paper are concluded as followed:(1)A multi-label inductive confidence prediction model based on CNN(MLICP-CNN)are proposed.The MLICP-CNN model included the inference rule extraction stages based on CNN using training set,and the multi-label inductive confidence prediction stage using validation set.In essence,MLICP-CNN model makes full use of CNN to extract sample multi-level and multi-angle characteristics automatically,and also uses the confidence mechanism and low time complexity of ICP to solve the problem of multi-label image confidence prediction,which makes the predicted results calibrated by the confidence.(2)Two nonconformity measure are designed to fulfill the MLICP-CNN learning framework,and thus two algorithms,IO-MLICP-CN and LS-MLICP-CNN,were proposed.A series of experiments on the standard multi-label datasets have been executed to demonstrate that the calibrated validity of confidence evaluation,domain prediction efficiency,and multi-label classification effect of the two models.(3)The better LS-MLICP-CNN model was applied to the prototype system for chest X-ray diagnosis for confidence diagnosis,which outputs multi-label diagnosis results attached with confidence evaluation.In addition,CNN were used to produce a pneumonia hotspot image to visually locate the effective lesion area.Finally,the work of this paper is summarized,and the prospects for future research are proposed.