Image Material Attribute Annotation Based On Boosting Algorithm And Stratified Gene Selection

Identifying material attribute based on visual characteristics is a hot research topic in the field of computer vision.Material type of an object's surface provides people a lot of valuable semantic information(softness,waterproofness,breathability,wearability,etc.),which is very important for accurately recognizing different materials.Therefore,the material attribute annotation research has very large practical value.This thesis focuses on image material attribute annotation.Several state-of-the-art technologies such as deep learning models,boosting algorithms,modified effective range based gene selection(ERGS)algorithm and stratified gene selection multi-feature boosting algorithm are used or designed for material attribute annotation.The main work is shown as follows:(1)Image material attribute annotation based on deep learning models: To allivate the problem of insufficient feature learning,the transfer learning methods byusing the pre-training models such as VGGNet,InceptionV3,DenseNet are used to train the annotation models: the low-level features are extracted firstly,then the target dataset is used for fine-tuning the annotation models,which helps to get the middle-level semantic information of images and complete material attribute annotation.Experimental resuls show that the state-of-the-art DenseNet model is superior to other models;the coarse-grained dataset obtains better annotation performance than the fine-grained dataset.The deep learning based annotation performance needs to be improved.(2)Image material attribute annotation based on boosting algorithms: Image features i.e.LBP,SIFT,Gist and deep learning based VGG16 feature are extracted respectively.Then the state-of-the-art boosting algorithms including XGBoost,GBDT,AdaBoost etc.are utilized to design the corresponding strong classifier by integrating a group of weaker classifiers.Based on the strong classifier,image attribute annotation is implemented.Experimental resuls show that the proposed boosting algorithm are better than the above mentioned deep learning models.The state-of-the-art XGBoost algorithm obtains the best annotation performance among all models/algorithms.It also means this algorithm is robust to different features and different datasets;the fine-grained dataset obtains better performance than the coarse-grained dataset;It is useful for matrial attribute annotation bytuning the number of weaker classifiers.Different single feature has different perspective about material attributes,which means single feature isn't enough for describing different material attributes objectively.(3)Image material attribute annotation based on a modified effective range based gene selection algorithm: the traditional effective range based gene selection(ERGS)algorithm is modified for material attribute annotation.It dynamically calculates the corresponding ERGS weight ofthe LBP,Gist,SIFT and VGG16 feature.And multi-feature fusion is completed after the weights assignments.Experimental results show that the modified ERGS algorithm makes full use of the complementarity between different features and annotation performance is further improved.The proposed GS-XGBoost model gets the best annotation performance.For the coarse-grained dataset,the "S(SIFT)+G(Gist)+L(LBP)” feature combination gets the best performance while for the fine-grained dataset,the "S(SIFT)+G(Gist)" feature combination is the best.The fine-grained dataset obtains better performance than the coarse-grained dataset,but its improvement is lower than that of the coarse-grained dataset.However,the proposed ERGS algorithm doesn't consider the effect of some stratified priori information during the material attribute annotation procedure.(4)Image material attribute annotation based on stratified gene selection multi-feature fusion: Stratified priori information(SPI)is obtained on basis of expert knowledge,historical data,andthe scatter figure of t-SNE.The SPI is absorbed into the modified ERGS algorithm and a novel stratified gene selection model is proposed.With the help of the stratified gene selection model,image material attribute annotation is implemented.Experimental results show that the stratified prior information is beneficial for material attribute annotation.To our surprise,very simple classification model can also obtain exciting classification results.For the coarse-grained dataset,the "S(SIFT)+V(VGG16)" feature combination gets the best performance while for the fine-grained dataset,the "S(SIFT)+G(Gist)+ L(LBP)" feature combination is the best.The fine-grained dataset gets better performance than the coarse-grained dataset,but its improvement is slightly lower than that of the coarse-grained dataset;"Average" pooling strategy plays an important role in the stratified gene selection multi-feature fusion procedure.