Research On SSD-based Method Of Detecting Typical Fittings In Transmission Line Images

The fittings is the power component which plays the role of fixation,protection and connection in the transmission line,corrosion,deformation,damage and other defects can easily occur in the transmission line fittings that has been in the harsh field environment for a long time,thus affecting the normal operation of the transmission line.The current transmission line inspection mainly relies on manual identification of power component defects in uav aerial survey line image,which not only consumes manpower and resources but also is inefficient.The present powerful capability of computer vision in object detection and recognition provides a new idea for automatic processing of inspection image,realizing automatic inspection of power components of transmission lines will greatly save manpower and material resources,and improve the efficiency of inspection operation and maintenance.This paper has mainly completed the following works on the detection of transmission line fittings based on SSD :In order to realize automatic detection of transmission line components,a transmission line fittings detection database containing various kinds of fittings was built,the dataset was expanded adaptively,and SSD was fine-tuned with this dataset.The fine-tuned SSD could basically complete the detection of the fittings object,but the model has poor detection effect in the case of small proportion of object in the inspection image and complex background of object.Aiming at the problem that the fine-tuned SSD model has poor detection effect in the case of dense targets,the following improvements were made to the model: in the matching stage of default box,the improved Io U was adopted to complete the screening of positive and negative samples,and the repulsion loss function was used in the model to replace the original loss function of SSD model.The improved SSD model has an average accuracy increase of 4.43% in 11 categories of fittings.In the fittings dataset,the marking boxes intersect widely.The intersecting mark boxes leads to the blurring of the category information of the fittings,which is not conducive to the detection of the fittings.In order to realize the embedding of the context information in the deep learning model according to the characteristics of the object,this paper completes the design of the occlusion relation module for the occlusion problem of the object on the basis of the existing relational module,and inserts it into the SSD model.In order to verify the effectiveness of the occlusion relation module,the index of occlusion degree was designed to quantitatively analyze the occlusion of the object boxes in the dataset of fittings.Finally,eight classes of fittings with high occlusion degree and sufficient number of marked boxes were selected for the experiment,the experiment proved that: in the test of the 8 kinds of fittings,the m AP of the original SSD model was72.10%,and the m AP of the SSD model embedded in the occlusion relation module was76.56%,with the performance improved by 4.46%.