Object Segmentation And Generation Method For Images Of Retail Shelves

Separating the semantic block information of different objects from the input image is an important research problem in the field of image segmentation and semantic recognition.The inverse problem,based on semantic information to generate realistic specified object images,is also an important task in the field of image generation.This thesis focuses on the practical application of image segmentation and generation of retail products.This thesis studies the reciprocal problems mentioned above and proposes an effective method to meet the needs of product object segmentation under diversified scenarios.At the same time,image data of single objects and multiple objects can be generated.Firstly,for objects segmentation,this thesis designs and implements an object segmentation method based on gradually relaxing the fixed background hypothesis to achieve accurate segmentation of objects in different scenarios.The method assumes a fixed background,uses the acquisition device and calculates prior graph to achieve automatic segmentation of the image.For scenes such as reflections that break the fixed background hypothesis,a multi-camera method is used to generate a visualhull and further projected to the image to form a visualhull constraint for reflection removal.In the case where a plurality of moving objects appear in a fixed scene at the same time,the deep learning method for image sequence segmentation is used to gradually learn the feature information of a specific target object from coarse to fine.In the qualitative and quantitative experiments,the user's tendency of over 60% and the average removal rate of over 90% indicate that the method can effectively segment the object images in different scene.Secondly,for product object image generation,an image generation method based on generative adversarial networks is proposed,which achieves the generation of a highquality realistic image with a small amount of guidance information.When generating an image of a single product,it is not necessary to constrain the category of the object.And an image is generated by random noise.When the product object category of the generated image needs to be specified,the guidance is assisted by the semantic information.When performing multi-item object image generation,use bounding box represented semantic label map to simplify the data annotation work and reduce the guidance information needed.In addition,the translator-enhancer framework allows the translator to focus on transforming semantic information into images from coarse-grained,while the enhancer to focus on the fine-grained detail of the image.Quantitative experiments show that the proposed method has a 3.04% and 1.04% improvement over the baseline method in terms of structural similarity index and peak signal-to-noise ratio.In the qualitative experiment,the user's tendency of over 50% also indicates the effectiveness of the image generation algorithm of the object.