Research On Deep Learning Based Product Family Imagery Design Method

With the continuous progress of science and technology and the continuous change of social needs,the personalized needs of users for branded products are receiving more and more attention from all walks of life and corresponding scientific research and technological updates are being carried out.At the same time,the feedback from the market has also raised the demand for rapid product iteration to meet consumer demand.The demand for product personalization can give the product a subjective aesthetic identity and bring a pleasant and enjoyable experience to the user,and scholars and institutions at home and abroad have conducted rich research on this.Product family imagery has the same aesthetic and design characteristics in a certain range,so that the family genes in the iterative process can be preserved.At this time,to use the product family imagery to fit or express the relationship between the user and the product family iteration,the traditional method of mathematical model extraction does not achieve a very accurate imagery recognition model for the identification of product shapes corresponding to product imagery.With the development of artificial intelligence in image recognition in recent years,key research results have been obtained in a wide range of fields through its powerful data learning capability and autonomous learning iterative feature.In this paper,we propose a deep learning and data mining based approach to generate sample product family imagery modeling,and the detailed task includes three stages: the first stage is to determine the product family imagery ontology model using the triad approach,so as to derive the three imagery dimensions of the product family.Then the relevant data is crawled according to these three dimensions,and finally the data is analyzed to obtain the product family imagery.In the second stage,we use the semantic difference method to select the imagery and modeling samples of different dimensions to form a questionnaire,and get the product family imagery modeling sample library,and use the improved residual network(Res Net)to identify the product family imagery to get the product family imagery identification model.This model is used to label the crawled samples to obtain the dataset.In the third stage,the target family imagery and the imagery of other dimensions are analyzed for lexical relatedness,and finally a matching dataset is obtained.This dataset is used as the database of the conditional adversarial network and the target family imagery is added as a condition,and the game mechanism of the generative adversarial network is iteratively trained to finally generate a product modeling sample design scheme with product family features and target imagery feature styles,which proves the feasibility of the model,and the imagery fit of the model is also verified in the next stage of expert evaluation.In this paper,the SUV series of Mazda product family and the feature models related to its family imagery are used as the research samples,and several imageries related to Mazda’s main imagery are generalized for them.These imagery product car models were established as a product imagery modeling dataset to prepare for the subsequent family imagery recognition experiments.After that,the Res Net network is used for parameter adjustment and model modification and the network is trained with the dataset to build the product family shape imagery recognition model.It is confirmed that Res Net is feasible for product modeling imagery recognition by its high recognition accuracy.Under the condition of imagery recognition.We propose a model for generating product family imagery modeling samples based on conditional generative adversarial networks.The iterative training samples close to the real samples are generated through the mutual game between the generator and the discriminator,and this scheme is able to generate product modeling sample design solutions based on the product family imagery features in large quantities,thus proving the feasibility of the model,and the imagery fit of the model is also verified in the next stage of expert evaluation.In this paper,we take advantage of the self-learning of convolutional neural network in deep learning to obtain image-compatible product modeling samples by training a large amount of data.It provides a new idea for product family iteration optimization and broadens the product family design method;it is able to incorporate the target imagery features into the new generation of product family solutions in sample generation,which in turn reduces the design cycle and labor cost,and thus improves the speed of product iteration in enterprises.