Research On Object Detection Technology Based On Deep Learning In Specific Scenes

In recent years,with the rapid development of deep learning and image processing technology,object detection has been one of the research hotspots in the field of computer vision.The task of object detection is to detect the interested targets in the image and determine the categories and locations.Because of the various appearance,size,pose,shooting angle and illumination condition of the object in the visual image,traditional object detection algorithm which depends on manual feature has great limitations.With the development of deep learning theory,the continuous innovation of algorithm and network structure,object detection algorithms based on deep learning has achieved excellent results on public dataset which provides a realistic and effective solution to object detection in many specific scenarios..Based on the deep learning theory,this thesis studies the object detection technology in different specific scenes.Including the cross domain object detection,the innovative Adaptive Region-based Fully Convolutional Network algorithm is proposed;for the face detection scene,the Focal Loss Multi-task Cascaded Convolutional Neural Network algorithm is proposed;for the actual industrial object detection,combined with the proposed algorithm and traditional image processing technology,the complete hardware and software system of bolt loosening and water seepage detection based on deep learning is built.The main work and innovations of this thesis are as follows:1.For the problem of cross domain object detection,there is a significant domain transformation between training data and test data in deep learning.Adaptive domain classifiers are introduced into the object detection algorithm,and an Adaptive Region-based Fully Convolutional Network algorithm is proposed.Based on the analysis and research of the principle and structure of the Region-based Fully Convolutional Network,the performance degradation caused by the domain transformation is considered from the two levels,image and object.The image level and object level domain classifiers are introduced into the network model as adaptive components,gradients of the domain classifiers are reversed in training stage,which endows the model parameters with adaptive property for cross domain object detection,and an Adaptive Region-based Fully Convolutional Network AR-FCN is proposed.Multiple open data sets are used in the different cross domain scenes,the experiment and analysis of the proposed innovative algorithm show that the Adaptive Region-based Fully Convolutional Network algorithm proposed in this thesis has better performance in object detection in various cross domain scenarios.2.For the common face detection in object detection,combining face detection with face alignment and considering the problem of class imbalance,Focal Loss Multi-task Cascaded Convolutional Neural Network algorithm is proposed.Based on the research and analysis of Multi-task Cascaded Convolutional Neural Network,considering the class imbalance in face detection,optimizing the multi-task loss function,Focal Loss Multi-task Cascaded Convolutional Neural Network FL-MTCNN algorithm is proposed,and online hard example mining and multi-scale training optimization are used in network training.The experimental results show that the proposed face detection algorithm has good performance in different complex scenes.3.Aiming at the problem of bolt looseness and water seepage detection in the industrial scene,a complete software and hardware detection system based on deep learning is designed and built by combining the proposed Adaptive Region-based Fully Convolutional Network algorithm with traditional image processing algorithm.For the problem that the bolts on the reservoir door are loose and water seepage,which leads to potential safety hazards,the whole detection algorithms can be divided into two steps.Firstly,the Adaptive Region-based Fully Convolutional Network algorithm is used to detect the bolt position and water seepage.Secondly,the traditional computer vision algorithms such as edge detection and Hough transform are used to calculate the straight line parameters of bolts,and the rotation angle is obtained by clustering algorithm.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.