Anomaly Object Detection Based On Deep Learning

Object detection has been one of the research hotspots in the field of computer vision since the advent of computer technology.The task of object detection is to find the objects in the image and determine the locations and categories.Traditional object detection algorithms relying on manual features often have no suitable solutions in environment where object pose and angle are variant.With the development of deep learning theory,new models and new network structures continue to emerge.Object detection algorithms based on deep learning achieve excellent results on public dataset which provides a realistic and effective solution to solve problems in actual engineering.Based on the deep learning theory,this thesis focuses on Region-Based Fully Convolutional Networks(R-FCN)and applies it to two scenarios: circular loom fabric texture detection and bolt looseness and water seepage detection.Complete software and hardware systems are built.The main work and innovations of this thesis are as follows:1.An object detection algorithm based on deep learning named R-FCN is studied and a global information fusion R-FCN network structure is proposed.Based on Faster R-CNN,R-FCN integrates a large number of independent operations by introducing position-sensitive score maps,which realizes a fully convolutional network and improves the detection speed while ensuring accuracy.The detection accuracy is improved by using the Soft-NMS algorithm in the sample post-processing of the R-FCN network.Only the local information block voting is used in R-FCN.This thesis proposes global information fusion R-FCN network which can take advantages of global information.Experimental results show that the global information fusion R-FCN network proposed in this thesis can achieve higher object detection accuracy with minimal loss of detection efficiency.2.To detect the defects in fabric,a Fast R-FCN network is proposed and a circular loom fabric texture detection system is built.In view of the four types of defects in hose spinning and the real-time requirements during detection,this thesis improves the R-FCN network by replacing ResNet-101 with ResNeXt-50.Considering the small size of the feature maps in strip fabric,the first downsampling in Conv5 layer is removed,and the dilated convolution is used to replace the ordinary convolution.The receptive field is increased,so that the output of each convolution contains a large range of information.Multi-scale alternate training strategy and Online Hard Example Mining are used to ensure the network is well trained.The experimental results show that the system constructed in this thesis is stable and reliable,meets the real-time detection requirements and can effectively detect the fabric defects on the strip.3.To detect the bolt looseness and water seepage,algorithms based on deep learning and Hough transform are proposed and a detection system is built.The reservoir doors have loose bolts and water seepage which will lead to safety hazards.In this thesis,the whole detection algorithms can be divided into two steps.Firstly,the bolts position and water seepage are detected by R-FCN in which the anchor sizes are changed.Then,the computer vision algorithms based on Hough transform are used to calculate the line parameters of the six sides of the bolt and the hierarchical clustering is used to judge whether the results meet the conditions for the formation of hexagon.The experimental results show that the system constructed in this thesis is stable and reliable and has high detection accuracy for bolt looseness and water seepage.