Super-resolution Reconstruction Of Spot Images Ablated By Femtosecond Laser On Silicon Crystal Based On Convolutional Neural Networks

In recent years,femtosecond laser has been more and more widely used in the field of noncontact,high-precision,micro-scale micro-nano processing.When machining microstructures,because it is not convenient to directly measure the depth information of the processing structure,the processing method of off-line measurement and adjusting related parameters is usually adopted.However,this method will introduce repeated positioning errors and make the depth information inaccurate.Therefore,it is of great significance to make full use of the information contained in the light spot image derived from femtosecond laser ablation of monocrystalline silicon to realize the closed-loop control of the depth of the groove processing in the future.The plasma emission generated by femtosecond laser during ablation of silicon crystal surface is in the form of laser point.In the process of ablation,the laser spot luminescence is very weak,and the phenomenon of melting and vaporization will occur on the surface of the silicon crystal,resulting in serious degradation of the core ablation area of the collected spot image.Aiming at the features of low contrast and signal-to-noise ratio in the edge and background of femtosecond laser plasma,as well as the core and halo of the spot,the methods to improve the image resolution of the original spot are studied.First,the collected plasma light spot image is taken as the original image,and it is enhanced by principal component analysis and bilateral filtering.Secondly,the enhanced image is stratified at the grayscale level.According to the obtained energy distribution information,the spot image is divided into spot core area,halo area and ablative transition area.Then,the three main areas are completely separated.Thirdly,the processed spot image is reconstructed with super-resolution based on convolutional neural network,and the high-resolution image is recovered from the low-resolution image.To improve and optimize the traditional CNN structure,and a deeper CNN structure is proposed,which finally achieves high-quality and rapid reconstruction of spot images.The experimental results show that the reconstructed spot image with the improved model is easier to recognize than the original image,and the edges and contours are clearer.Moreover,the PSNR value obtained by the improved algorithm is 1.868 d B higher on average than that of the traditional algorithm,and the reconstruction time is only 30% of the traditional method.Figure 36;Table 2;Reference 52...