Image-based Analytic Strategy With Neural Network For Product Semantics

Design is an important link in the strategy of manufacturing power,and the shortage of product design ability is the bottleneck that affects the upgrading of manufacturing industry in China.Product design originates from obtaining user demand.With the increasing competition in the market,users’ demand for products shows a diversified and personalized trend.Product utility is no longer the most important factor for consumers to buy,and the demand for product sensitivity is being paid more and more attention to.Image is an important medium for users and designers to communicate sentimental demand.However,due to the diversity and polysemy of image information,and the perception difference between designers and users,it is difficult for designers to accurately understand users’ sentimental demands.Since semantic information of image is a complex nonlinear mapping problem,there is a lack of models,methods and tools to help designers understand it accurately and efficiently.Therefore,it is urgent to study the Image-based analytic strategy with neural network for product semantics,establish the product semantic intelligent resolution tool,and help the designer understand the product semantics of the user.Based on the project of the National Key Research and Development Program(No.2018YFB1700702)and Sichuan Province Major Science and Technology Project(No.2019YFG0397),this paper aims to implement intelligent and accurate resolution of product semantics and assist designers in understanding user’s sentimental demand.The main research contents of the thesis are as follows:(1)In Chapter 1.This paper discusses the research background of product perceptual semantic intelligent evaluation model based on visual guidance.This paper mainly analyzes the current research situation in the related fields of emotional design,Kansei engineering,product emotional image semantic evaluation and product intelligent design.Then,this paper analyzes the shortcoming of perceptual vocabulary selection,similarity calculation of product semantics,perceptual identification of product details and perceptual evaluation of product.At last,the paper puts forward the research objective and content.(2)In Chapter 2.The key techniques of user perception need are studied to help designers understand,such as product semantic lexicon selection method,product perception similarity calculation method,image data enhancement algorithm and image classification algorithm.Then the main process and general framework of image-based analytic strategy with neural network for product semantics are presented.(3)In Chapter 3.In order to solve the time consuming and uncertainty,a product semantic word screening method PCA-E(Principal Component Analysis-Exploration)is proposed.We first constructed the user evaluation matrix for the SD questionnaire.Then we used PCA method to reduce the dimensions of the matrix,and finally used PCA-E explanatory principle to explain the specific significance of each evaluation dimension by relying on component load coefficient.At last,This method is compared with the traditional method through the perceptual word selection experiment of automobile products.(4)In Chapter 4.Since there is a gap between the evaluation results of ideal recommendation models and actual user preferences in recommendation accuracy,we propose a PSIM(Product Perceptive Similarity Model)algorithm to calculate product similarity.We establish a product similarity cognition process based on knowledge level.On this basis,the system diagram of PSIM is determined.The emotional preference,conceptual conflict and conceptual judgment of the product are calculated by using the intensity comparison,contrast comparison and relatedness comparison of the user evaluation matrix.So that the perceptual similarity of the product can be used for reasonable recommendation.By applying this method,the similarity recommendation of three kinds of mass market products,automobile,clothing and shoes,is completed.(5)In Chapter 5.In order to solve the problem of image overfitting caused by focusing on the whole image and ignoring the parts in machine learning,Mesh Cut algorithm is proposed.During the training stage of neural network,the algorithm converts an image into a mosaic of multiple image segments by superimposing a stripe mask on the image.This method decomposes the whole feature of the target into several local features and provides multi-angle information for the network.Finally,we experimentally validate the image classification,target detection,and semantic segmentation of computer vision tasks,and analyze the effects of each parameter adjustment in Mesh Cut on the results.(6)In Chapter 6.Due to the problem of nonlinear mapping and strong generalization model in product perceptual attribute evaluation,a new method for fuzzy evaluation of perceptual attribute of deep learning assistant,DLFAE(Deep-learning-assisted Fuzzy Attribute-Evaluation),is proposed.This method combines subjective evaluation with computer technology.Firstly,this method uses the network to train the small sample of emotion attribute of product image calibrated by SD questionnaire.Then the fuzzy evaluation model of product perceptual attribute is established.Finally,the model is used to train the uncalibrated samples to generate the evaluation results automatically.The results are compared with Res Net-18 and VGG-16 in the men’s cloth perceptual attribute classification experiment.(7)In Chapter 7.Combining with the key techniques mentioned above,a Image-based analytic strategy with neural network for product semantics tool is developed.This tool contains several modules,such as user basic information collection,photo upload,product sensitivity automatic analysis,similarity picture recommendation,etc..We simulate the vehicle and elevator products as experimental data.