Fabric Defect Detection Based on FCN Uncertainty Feature

In order to improve the accuracy of fabric defect detection based on Full Convolutional neural network (FCN) algorithm, a fabric defect detection method based on FCN uncertain features was proposed. Based on the model features of the stacked denoising automatic encoder, the natural attribute of detection uncertainty is added to the detection system, forming a new hybrid FCN model. The over fitting of the model will cause the performance of the training samples to be too superior, resulting in a large Generalization error of the model. In order to prevent the FCN over fitting phenomenon and improve the detection efficiency of the network in the network training process, this paper introduces Dropout in the full connection layer, and temporarily discards the neural network unit from the network according to a certain probability. Image artifacts seriously affect the quality of images. A new hybrid downsampling method is proposed to reduce image artifacts caused by the convolution process through deconvolution algorithms. Experiments have shown that this algorithm can effectively improve the accuracy of FCN foreign object detection, and the PR curve of the algorithm is superior to existing algorithms, which can achieve effective detection of fabric defects.