Research On Methods Of Tower Number Plate Detection In Transmission Line Images Based On Convolutional Neural Network

Tower number plate detection refers to extracting areas of the tower identification plates from the transmission line scene images,which provides technical support for subsequent plate number recognition,defect detection,and database development.With the continuous development of machine learning technology,deep learning-based object detection method has become the mainstream.However,for the real-world line inspection images taken by drones,which typically have large view fields and relatively small number plate areas,the tower number plate detection has low accuracy,slow processing speed,unable to meet application needs.Therefore,this paper proposes the project of research on methods of tower number plate detection in transmission line images based on convolutional neural network,working on tower number plate detection of wide field of view images from three aspects: cascade detection structure,data augmentation,and multi-scale feature enhancement,aiming to improve the accuracy and speed of inspection line tower number plate detection,which has very important practical value.The main research work of this article is as follows:(1)Based on the object detection network FPN,we carry out a detailed comparison and analysis of two wide fields of view tower plate detection methods: direct full image detection and block-wise detection.Among them,the former is carried out mainly by first resizing the entire large image and then performing number plate detection,when the computing resource is limited.The detection accuracy is low since the scales of the number plate areas are relatively small.The latter is mainly to detect in blocks of wide-field images,resulting in disadvantages of high false detection rate and slow speed.Therefore,the research focus of this paper is to improve the detection accuracy and speed of small-scale tower number plates.(2)To address the problem of small-scale number plate areas relative to the entire inspection image,a coarse-to-fine cascade detection method is proposed by taking advantage of the contextual relationship between number plates and relatively large-scale tower.First,detect the tower area on the resized low-resolution image;then,split the tower area into multiple overlapping sub-areas as the search areas for number plates according to the morphological characteristics of the tower area;and finally,extract the number plate areas in the sub-areas of the corresponding high-resolution image,which can improve the accuracy of the number plate detection.Besides,for the problem of insufficient number plate samples,an ensembled data augmentation strategy combining image samples expansion,number plate pasting,and training images stitching is proposed,which increases the number and diversity of training samples.The experimental result shows that,compared with direct full-image detection,the precision of the proposed cascade detection method is improved by 18.2%.(3)To address the problem of insufficient feature representation of small-scale number plates under complex backgrounds,a feature network with attention-based enhancement and multi-scale fusion is proposed.First,the adjacent feature layers of FPN’s feature pyramid are fused to obtain enhanced feature maps;then,a non-local-based network combing channel-wise and spatial-wise attention blocks is used to enhance object areas in the feature maps.The experimental results show that the proposed feature enhancement module can improve the representation ability of the FPN feature network,especially for small-scale number plates,which can relatively enhance the target features while suppressing the background.Compared with FPN,the detection precision is improved by5.1%.Combining with the cascade detection and training samples augmentation methods,compared with the block-wise detection algorithm,the precision of the proposed algorithm is improved by 0.3%,the false detection rate is reduced by 37.4%,and the detection speed is increased by 10 times.This paper proposes a cascade detection and attention enhanced multi-scale feature fusion method.Experiments show that the proposed method can improve the precision and speed of tower number plate detection,providing a new technical approach and key technical support for transmission line tower number plate detection and identification.