Research On The Application Of Adaptive Bilateral Filtering Threshold Segmentation Algorithm In Fabric Defect Detection

In order to better segment the fabric defect image and improve the accuracy of fabric defect identification,the related research of fabric defect is mainly carried out through three aspects.The first aspect proposes an adaptive bilateral filtering algorithm based on the Stable Mean Difference Criterion(SMDC),which mainly improves the disadvantages of traditional bilateral filtering that need to be adjusted manually based on experience.The algorithm mainly uses SMDC to achieve the main parameters Adaptive adjustment.For each fabric defect image,there are suitable parameters corresponding to it.Experimental results show that the bilateral filtering algorithm tends to be stable during the automatic adjustment of the parameters,the SSIM reaches the adaptive value,and the fabric defect image segmentation results are clear;In the second aspect,an improved Otsu segmentation algorithm is proposed,which mainly has a good segmentation effect on the bimodal type images of the traditional Otsu segmentation algorithm,but improves the problem of the unimodal type of fabric defect images.The algorithm mainly adjusts the threshold by calculating the similarity between the segmented defect image and the original defect image to enhance the Otsu segmentation algorithm's self-adaptation ability for single-peak type images.The experimental results show that the improved Otsu segmentation algorithm can not only achieve a better segmentation effect when the histogram of the fabric defect image is single peak,but also retain a good segmentation effect when the histogram of the fabric defect image is bimodal.,In line with the expected results of the experiment.The third aspect mainly studies the application of neural network technology in fabric defect detection and recognition.The neural network model densenet169 is mainly used,and the loss function is LSR(Label smoothing Regularization),which reduces the over-fitting problem during the model training process.The super selection is 0.25 and 0.3.The experimental results show that the neural network model predicts the results more accurately.