Research On Digital Print Defect Detection Algorithm Based On Image Recognition

Digital printing quality detection is the key to industrial development,and its quality plays a very important role in people’s life.Common methods in the field of digital printing image defect detection include artificial vision detection based on traditional methods and deep learning technology detection.Among them,artificial vision detection requires a lot of time and personnel,which not only increases labor costs,but also fails to meet the demand of high speed and high accuracy of industrial defect detection.At present,image processing technology develops rapidly and has gradually become an effective means of defect detection.Based on image processing technology,this paper carries out the following research:(1)Pre-processing part: Based on the defects of complex information and difficult processing of digital printing image,the weighted average method is firstly used for graying,then the histogram equalization is used to enhance the gray image,and finally the adaptive median filter is used to achieve image denoising.(2)Defect image edge detection part: Expand the Sobel operator,Roberts operator,Laplace operator and Canny operator are compared,the Canny operator is improved and applied to the digital printing data set for verification.The adaptive median filter is used to replace the original Gaussian filter,and 45 and 135 are added to the horizontal and vertical directions of the original Canny operator.The gradient amplitudes were obtained from two directions.Finally,aiming at the defect that the threshold value of the original Canny operator needs to be adjusted manually,the two-dimensional maximum entropy method which can adjust the threshold adaptively was proposed through experiments.Compared with the traditional algorithm,the average structural similarity edge preserving index and the average pixel value of the improved Canny operator method were improved.(3)Defect classification: Based on the transfer learning method,Dense Net169 neural network is combined with transfer learning to obtain the training weight,and CBAM attention mechanism module is added to the Dense Net169 network model for better feature extraction.Finally,the improved model is compared with some early deep learning networks.The accuracy rate reached 88.8%,4.1% higher than the traditional Dense Net169 network,and the Kappa coefficient reached 0.857,5.41% higher than the traditional Dense Net169 network.