Casting Defects Image Generation Research Based On Semantic Description

With the development of computee reduced the gap betwr eevisnio mn anan ad ndar mtificial intelligence technology, semantictechnology havachine interaction. How to establish ahuman computer interaction system based on high-level semantics to generatesnon-parametric images is the current technical challenges. This paper aims to provideeffective analysis, implementation and evaluation methods for non-parametric defect imagegeneration and semantic–driven. Ultimately implement an available non-parametric defectsimage generation algorithm based on semantics–driven. According to the semantic-drivenrequirements and characteristics of defect images, the main issues involved in the proposedalgorithm include: parsing semantic input accurately, eliminating the semantic gap, diversityimage generation.Firstly, this dissertation analyze image from the perspective of image features, summedup in the description of image shape features have the characteristic of intuitive, goodstability, and possess semantic information, etc. Then introduce several typical shapedescription algorithm. And learned that geometric parameters such as corner, curveparameters is closer to the human visual perception system, which laid the theoreticalfoundation for the framework described below.Secondly, this dissertation analyze the defect structures from human visual point, andpropose a defect skeleton generated algorithm based on the angle and length of the curve.Then design experiments to inject semantic information for the skeleton and build a smallsemantic vocabulary.Thirdly, the process of defects simulation is divided into five stages:weight assignment,skeleton proliferation, resizing, adding gray, edge processing. In the weight assignment stage,this dissertation assign different weights according to the skeleton structure of geometricimperfections. In the skeleton proliferation stage, the algorithm simulates dendritic shrinkageforming process and propose a framework structure evolve over time and controlled diffusionalgorithm based on Gaussian filter. To make the simulation more similar to real defects anddeficiencies, algorithm add the gray information according to the real defects and introducethree parameters such as defect area, diameter, and the local density to revise the size of thedefect. At last, blur the edge to make the defect blend with the background more natural. Finally, this dissertation show the experimental results of the algorithm. Take advantageof the defect detection algorithms to detect the simulated and real defects and give anobjective evaluation. At last, summed up the deficiencies of simulation algorithm and thefuture direction of the algorithm.