Research On Defect Image Generation And Detection Based On Deep Learning

Mining is a high-risk industry,and ensuring its sustainable development requires a high degree of attention to safety production.The detection of defects in mining shaft facilities is an essential part of safety production.The use of monitoring videos of shaft facilities and methods based on image depth feature learning can effectively solve problems such as high cost and safety threats associated with manual detection.However,when the number of samples of defective facilities’ images is insufficient or the samples of defects are incomplete,the feature expression ability of shaft facility images may be low,which may lead to low accuracy in defect recognition.In this paper,we propose an improved Defect Cycle-Consistent Adversarial Network(CycleGAN-De)to generate mining shaft facility defect image data and develop a hidden danger detection method of mineshaft facilities based on deep image feature learning(HDD-DIFL).The HDD-DIFL method learns a defect recognition model based on high expression features of normal image samples to improve the accuracy of defect detection in shaft facilities.The research in this paper is mainly carried out in the following two aspects:(1)We propose an improved CycleGAN-De for generating realistic defect image samples.First,we construct an adversarial generative network structure that includes two generators and two discriminators.One generator is used to convert normal images into defect images,and the other is used for reverse conversion.To improve the quality of generated images,we use the Dense Block module of the multi-level residual network in the generators.Then,we design a joint loss function that includes cycle consistency loss and adversarial loss to train the generative network.Finally,we use the trained model to expand the defect image dataset.Experimental results show that compared with classical advanced generative models,the proposed model can generate better quality shaft defect images,providing a necessary dataset for subsequent shaft facility defect detection.(2)Due to the incomplete defect image dataset and the scarcity of defect images,we propose the HDD-DIFL algorithm to study a single-class defect detection model based on a normal image dataset.We first construct a dataset containing a large number of normal facility images and a small number of defect images.Then,we use a parallel neural network consisting of a feature extraction network and a classification network as the training network.Next,we design a joint loss function to train the network on normal image sets,external image sets,and a small number of defect image sets.Finally,we use a clustering strategy to select representative normal image depth features as feature templates and use template matching to implement defect image detection.Experimental results show that on the constructed mining shaft interior scaffolding facility image dataset,the feature extraction model trained by the proposed algorithm can effectively improve the expression ability of image features,and the defect detection accuracy of the model can reach above 92%,providing a new idea for using mining shaft interior facility images for defect detection.