Study On The Classification Of Femtosecond Laser Spot Images Based On Manifold Learning And ELM

Femtosecond laser processing is a high-precision micro-nano processing method,but the real-time detection of the ablation depth of the object to be processed is a technical problem in the processing.It has been proved that the single-layer processing depth of microstructure is determined by laser ablation power,processing speed and feed rate.The determination of ablation power can estimate the processing depth to some extent.When silicon material is ablated by femtosecond laser,plasma derived luminescence also occurs.In order to explore the relationship between plasma spot characteristics and ablation power,the feature extraction and analysis of the spot image can be used to realize the classification and recognition of femtosecond laser plasma spot images under different ablation powers.Considering the low signal-to-noise ratio of plasma spot image and the inconspicuous contrast between the edge and the background area,the selected color space hybrid filtering combined with the principal component enhancement method is used to obtain the spot image with strong contrast and effective area,and the spot tail dissipating part is filtered out.The obtained first principal component is subjected to pseudo color processing to analyze the energy distribution.The desired region is obtained according to the energy distribution.The segmentation of principal component extraction is used to obtain the complete spot image,and the geometric features of the region are analyzed.It is found that the geometric characteristics with the same ablation power have strong consistency.The core ablation zone is obtained by threshold segmentation,and its centroid and long and short axis are extracted for judging the processing direction.The luminance mask component of the binary mask obtained by Niblack segmentation obtains effective luminance information and extracts luminance features.It is found that the total brightness of the spot has strong stability,and the correlation coefficient with the ablation power is greater than 0.8,which can be used for the identification of laser ablation power.The six-dimensional feature matrix is composed of the geometric features and brightness features.The clustering effect is analyzed by the manifold learning and dimensionality reduction.The good clustering characteristics indicate the validity of the selected features and the feasibility of implementing the classification.Finally,a classification model of popular learning combined with extreme learning machine can be built,which can achieve 97.3% classification accuracy in 0.0027 seconds.Figure 32;Table 4;Reference 54...