Research And Implementation Of Multi-Spectral Feature Fusion Optimization Segmentation Algorithms Based On Memory Gene Model

The excellent traditional Chinese culture created and continued by the Chinese nation in thousands of years of history is the root and soul of the Chinese nation.Traditional patterns are an important part of Chinese traditional culture.It is necessary to promote the expression and excavation of Chinese traditional culture through research in the field of patterns.This thesis takes traditional patterns as the main research object,and based on the meta-learning method,combines the phase spectrum features and amplitude spectrum features of patterns to carry out semantic segmentation research on patterns,and finally builds a traditional pattern segmentation system.The main research contents include:(1)Construct a traditional pattern data set.Through laboratory accumulation,scanner scanning,museum open resources,etc.,the original pattern data and material data are obtained.After preprocessing the collected pattern images,pixel-level annotation is performed to provide data support for subsequent data set expansion.(2)Propose an optimization strategy for multispectral fusion and introduce it into the semantic segmentation algorithm.Perform Fourier transform on the pattern image and material image to obtain two sets of phase spectrum and amplitude spectrum.After exchanging the central part of the amplitude spectrum,the reconstructed pattern image is obtained through inverse Fourier transform.As an extended data set,the reconstructed pattern images can introducing the phase spectrum features into the model.Experiments prove that the semantic segmentation results can be optimized by comprehensively considering the magnitude spectrum features and phase spectrum features.(3)Propose a small-sample semantic segmentation algorithm based on the memory gene model.Firstly,a memory gene model is proposed to guide the algorithm optimization ideas.The meta-learning model is used as the network framework to initialize the model parameters.Acquire prior knowledge through the saliency detection module.And through the multiscale feature fusion module,further learn and retain common knowledge.Experiments have proved that based on the above method,the segmentation accuracy has been improved to a certain extent compared with other algorithms.(4)Design and implement the traditional pattern semantic segmentation system.After analyzing the requirements of the system,the system is divided into four parts:registration and login module,semantic segmentation module,reconstructed image demonstration module and material library module.The function design and implementation form of the modules are introduced in turn,and finally the system test is carried out to ensure the integrity of the system.stability and usability.To sum up,this thesis proposes a semantic segmentation algorithm applied to traditional patterns,and based on this algorithm,realizes the semantic segmentation system of traditional patterns,and achieves good results on pattern datasets.