China is rich in bamboo resources.Due to its good physical and chemical properties and environmental friendliness,bamboo is widely used in engineering structures such as construction industry,mechanical and electrical product packaging industry and so on.Due to the inherent porous structure in the growth process of bamboo and the gluing process of engineering bamboo materials,the main failure form of engineering bamboo is the material fracture caused by the expansion of initial crack,which can not be ignored.The crack propagation length has always been one of the difficulties in the research of composite fracture.Using digital image correlation(DIC)technology to detect the crack tip in engineering bamboo fracture experiment can accurately calculate the length of crack propagation and provide effective support for the reliability calculation of fracture strain energy.However,this technology has high requirements for image quality,and there are many limitations in improving image quality from the hardware level.At present,in terms of software,the traditional interpolation method is still used in the reconstruction of engineering bamboo speckle image facing DIC technology,and the performance is relatively backward.Therefore,in order to accurately calculate the crack propagation length of engineering bamboo in mechanical experiment,based on sup er-resolution technology and deep learning technology,this paper studies the super-resolution reconstruction method of this kind of image around the speckle image of engineering bamboo,and mainly completes the following research contents:(1)According to the basic principle of DIC technology,the image acquisition system of engineering bamboo mechanical property experiment based on DIC technology was built and the related model was analyzed.Sprayed black-and-white paint on the engineering bamboo specimen to form speckle,drew lines,located and conducted friction treatment,installed it on the universal mechanical testing machine,gradually load ed it until the crack expanded obviously or the load changed suddenly,collected the digital speckle image in the loading process at the same time,form ed the engineering bamboo speckle image data set,analyzed the problems existing in the measurement of speckle image crack length,and analyzed the corresponding model.(2)Aiming at the problems of insufficient stability of network training and difficulty in extracting deep information from images,an attention-intensive residual network model for DIC images of engineering bamboo was proposed.The BN layer in the residual network was removed to improve the stability of the network and reduce the computational complexity,and the1 loss function was used to supervise the model training.Attention model was introduced to integrate the idea of dense connection into network structure.The attention-intensive residual block was designed to achieve the integration of information at different levels,effectively deepening the network depth and improving the network performance.The objective evaluation indexes PSNR,SSIM and subjective evaluation index MOS were used to evaluate the model performance.(3)Aiming at the problem that the edge of the reconstructed image was not sharp enough and the details were not rich enough,a generative adversarial network model based on attention-intensive residual structure and relative mean was proposed.The deep dense residual network with attention mechanism was used as the generating network,and the discriminant was improved by referring to the relative mean generating confrontation network.The training was supervised by combining the perception loss,content loss and confrontation loss,and the network interpolation was used to reduce the artifacts while maintaining the texture.A more reasonable image super-resolution model was constructed and applied to the engineering bamboo speckle image.The image quality and crack accuracy restored by different algorithms were compared and analyzed.(4)The super-resolution reconstruction system for engineering bamboo speckle image was realized.The main functions included user registration and login,image model training and image reconstruction test.The visual interface could freely set parameters,select paths and models. |