Product Image Attribute Annotation Based On Feature Learning And Effective Range Based Gene Selection Algorithm

Recently,with the rapid development of e-commerce,the number of product images on the Internet increase dramatically every day.An important prerequisite for effective management of the large-scale images is to make them "clear",which means we should more accurately and more effectively annotate the product images.As we all know,annotation method include manual annotation and automatic image annotation(AIA).The labor cost of manual annotation is so high,and there is little objectivity in the annotation results.Moreover,the early automatic annotation of product images mainly labeled the product images with nouns,the annotation results usually were unitary and contained many noises and ambiguities.The use of high-level composite semantics to annotate images(such as sentence annotation or phrase annotation)requires a large number of natural language processing algorithms to analyze the grammatical relationship between the words.The algorithm is very complex,which may increase the complexity of the annotation model.Rather than nouns annotation or sentence annotation,the thesis focuses on product image attribute annotation which is a new kind of annotation mechanism and lies on the middle semantic layer.The main works of the thesis are descrbed as the following:First,it is proposed that product image material attributes annotation model based on the traditional features and classification method.Material is a kind of distinctive feature to portray the product images.Its quality will affect users' purchase behavior in some extent.The thesis establishes a new material dataset named “MattrSet”.Based on the “MattrSet”,the LBP,Gist and SIFT features of the product images are extracted respectively from the perspectives of shape,texture,etc.The traditional image classification models such as KNN and Na?ve Bayes are used in turn to annotate the material attributes on the product images.In addition,a new transfer learning strategy is performed accross heterogeneous products to further enhance the annotation performance.Different from the traditional nouns annotation,the attribute annotation is a kind of adjective annotation.It realizes the annotation procedure across different product type and provides more and more semantic information for users.Experimental results show that the material classification performances of different image features are different.The overall classification performance of each image feature is very low.This means the description ability of each image feature has its own emphasis.Obviously it is very hard to describe the product images only by a single image feature.We need fuse different image features to describe the product images more comprehensively and improve the annotation performance.Second,product image attribute annotation based on deep learning features and ERGS(Effective Range Based Gene Selection,referred to as ERGS)algorithm is proposed.Deep learning features like VGG-16,VGG-19 are extracted to describe the product images.And ERGS algorithm is introduced to dynamically calculate feature weights and achieve feature late-fusion.Annotation models with better discriminative performance are generated in turn,which expands the deep semantic description of material attributes and enrich semantic meaning of material annotation.In addition,the annotation performance is improved by using the transfer learning strategy.Experimental results show that: 1)Annotation performance is improved significantly by using the ERGS algorithm;2)The deep-level semantic descriptions contain more valuable information(they are also called practical attributes),which can narrow the “semantic gap” between high-level human cognitions and low-level image features;3)The transfer learning strategy can be better fused into the annotation mechanism to complete annotations across different product categories,and the annotation performance has steadily improved.Third,a new relative attributes annotation model of product images is proposed because the practical attributes are relevant to the objective cognition of human being.Based on the Relative Attribute(RA)model,the practical attributes are annotated on the product images: unlike the traditional image attribute annotation(It is a kind of binary annotation.The annotation results are "existence or not").However,RA model makes a quantitative measurement of the practical attributes to better compare different product images.The new annotation results lie on the middle semantic layer,which can better assist the user's purchase behaviors.Experimental results show the annotation results obtained by the relative attribute model,which use zero-shot learning or a just small amount of sample learning,are superior to the traditional binary attribute annotation.Therefore,the practicability of the annotation model is greatly improved.The main innovations of the thesis are: 1)It focuses on the material attribute annotation and establishes a new material attribute dataset named “MattrSet”.Based on the ERGS algorithm and the RA model,it constructs a novel feature late-fusion model and complete product image annotation from the new perspective of attributes.2)It designs several evaluation methods for product material attribute annotation from the perspectives of features,materials,and kernel functions.They contribute to more objectively and more comrehensively evaluate our models.