Substation Object Detection Based On Context Information

With the popularization of the unattended substation,the research on the intelligent detection method of equipment based on substation patrol images has great application value in equipment monitoring,abnormal prewarning and other aspects,and is of great significance to ensure the safe operation of substations.To meet the need of equipment detection in the substation,this thesis uses deep learning technology and digs into the scale context information and spatial context information in substation images,to carry out the research on substation object detection method based on context information.The main research contents and innovative work of this thesis are as follows.Due to the diversity of equipment scales,the special scale objects are easy to be missed,resulting in a low recall.Aiming at this problem,this thesis proposes a method of substation image object detection based on scale context by using the prior knowledge of equipment scale at the data level and the multi-scale convolution information at the feature level.At the data level,on one hand,aiming at the problem that the camera angle of patrol inspection is limited and the multi-scale image of equipment is difficult to obtain,this thesis adopts the data enhancement method based on affine transformation to enrich the scale information in the training samples.On the other hand,aiming at the problem that the preset anchor scale is difficult to adapt to the scale distribution of substation equipment,this thesis proposes a scale formulation strategy of anchors,which uses the dataset scale prior knowledge to modify the anchor size and height-width ratio,to optimize the extraction of region proposal.At the feature level,A multi-scale feature fusion detection model is proposed to solve this problem,that the convolutional neural network can extract features in layers but the single-scale feature expression is insufficient.By fusing the information of the object at different scale feature layers,the loss of feature details at the high level is alleviated.The experimental results show that the proposed method can reduce the missed,and has the advantage of high recall.In substation,there is a strong spatial correlation between the object and the environment or between the object and the object,while the general object detection model often ignores the spatial context,resulting in inaccurate positioning or even unrecognized of some objects.Aiming at this problem,this thesis proposes an equipment detection model based on spatial context by using the spatial correlation of related environment and related objects.According to the characteristics of the relatively fixed scene settings and the co-presence of multiple objects in the substation image,this thesis models the spatial context,and designs two schemes of extracting full-image features directly and expanding the region-of-interest to extract the global context from the related environment.From two aspects of the visual clues and spatial position relations,the influence of other objects on the current object is integrated,and the local context from related objects is extracted.Then,the cyclic recursion and longterm memory characteristics of the recurrent neural network are used to input spatial context.Finally,the two kinds of spatial context are merged by mean-pooling,and the feature of the object is updated.The experimental results show that the proposed model effectively utilizes the spatial context and improves the accuracy of recognition and positioning on the substation images.