Research On Laser Coloring Technology Based On Deep Learning

Laser coloring has a wide range of applications in materials such as anti-counterfeiting,marking,and information storage.To achieve the best laser coloring effect,there are still many technical problems need to be addressed.For example,metal laser coloring depends on a lot of laser parameters,making it difficult to grasp the law of color changes with laser parameters.Therefore,a laser coloring method based on deep learning was proposed,and the main research contents are as follows:(1)The principle of laser coloring was analyzed and an experimental method of nanosecond laser coloring on the surface of stainless steel was proposed.A laser marking platform including a nanosecond pulsed laser with a wavelength of 1064 nm,a galvanometer system and a three-dimensional motion stage had been built,based on which a series of experiments were carried out.The whole experimental data were analyzed by Python,and the laser coloring pictures were subjected to projection transformation and vectorization processing.With two neural networks built in the Tensor Flow framework,the laser processing parameters were predicted and inverse design was also achieved.(2)The effect of laser processing parameters on the coloring effect was experimentally studied.The effects of laser processing parameters including average laser power(4 ～ 20 W),pulse width(2 ～ 350 ns),repetition frequency(20 ～ 1000 k Hz),scanning speed(10 ～ 740mm/s),scanning interval(0.001 ～ 0.03 mm)and defocus(0 ～ 3 mm)on the coloring effect of stainless steel was studied.After analyzing 1620 groups of effective laser coloring data,a conclusion was proposed that the laser pulse width and repetition frequency had little effect on the color,and the colors caused by changing the average power,scanning speed,scanning interval,or defocus were similar,which changed in sequence among orange-red,red,purple,blue,and green.(3)Two neural networks had been built,and laser coloring pictures had been processed by Python for shape correction and vectorization.A forward prediction neural network composed of 10 fully connected layers had been built,with a color predicting accuracy reaching to 95%,and a predicted chromatic aberration less than 7.Simultaneously,a backward prediction neural network composed of 8 fully connected layers had been built,with the trained forward prediction network connected in series.The corresponding laser processing parameters predicted from the colors were further verified by the forward prediction network,with an accuracy of 98%.A program was designed for the contour and color reading of the bitmap photos taken from the reality two-dimensional patterns,and the perspective transformation formula was employed for shape correction.Based on the Potrace algorithm,the reshaped bitmaps were vectorized.The research results of this thesis can greatly reduce the research and development time and material cost of the metal laser coloring process,and have guiding significance for practical industrial applications.