Based On Convolutional Neural Network Research On Infrared Image Detection Of Electrical Equipment

Object detection algorithms aim to detect the positions and the categories of the objects in images.These algorithms have been used in various areas in recent years.These algorithms can automatically classify equipment and locate abnormal temperature areas in facilities.Therefore,it is very important to implement object detection algorithms on infrared images collected by infrared thermal imager in power equipment automatic routing inspection systems.This method has been treated as an emerging technology,thus become more popular in power industry.Detection errors and misidentification are usually caused by the constant change of the devices' contours,as well as the emerging of a number of redundant candidate anchor boxes.These contour transforms are generally caused by changes of camera's shooting angle while the infra-image collecting period.Aim to solve these problems,in this thesis,we improve Cascade RCNN networks,and obtain multiple outcomes.First,we add the deformable convolution layer to the feature extraction network,and use the small offset to adapt to the changes of the image.Moreover,we adopt the probability score map of pixels to guide generate the center point of the location candidate area,thus reduce the numbers of candidate anchor boxes,and gain more meaningful suggestion candidate boxes.Finally,we used loss functions to supervise learning process,and improve the imbalance of positive and negative samples by focal loss function in the generate period of the center position.We used public data set and self-build ELE data set to train and test our algorithm.The experimental results show that our propose method,the Deformable Convolutional Guided-Cascade RCNN algorithm,can significantly improve the accuracy of infrared image detection of electrical equipment.The abnormal temperature regions of electrical equipments in thermal images are often small.Classical algorithms do not work efficiently on small target regions finding.In this thesis,a multi-scale feature extraction algorithm based on fusion equalization is proposed,which is further improved aforementioned algorithm.After feature extraction by Res Net,feature pyramid network(FPN)is used to combine low-level and high-level semantic information.The global semantic information is used to refine the features.The experimental results show that the fusion equalization multi-scale algorithm is effective in detecting anomal temperature areas of small targets.It can also accurately detect those areas in low contrast ratio.All the object detection algorithms based on deep learning method require a large data set for training.However,the resources of infrared images of electrical equipment are extremely scarce currently.On the basis of infrared images generated during the inspection of 110 KV and 35 KV substations in 7 regions of Tianjin provided by a power supply company in Tianjin,we established and labeled a data set called ELE,which contains 9770 thermal images and abnormal temperature areas of 6kinds of power equipment.These equipment covers insulator,current transformer,arrester,circuit breaker,disconnector and bushing.The ELE data set provides plenty of samples for training and testing,which effectively prevents overfitting.