Multi-Size Target Detection For Transmission Line Based On Deep Convolutional Networks

As the development of smart grid construction,target detection and recognition has become a new important technique for transmission line inspection.However,due to the scale diversity between targets of the transmission line like hammer,spacer,insulator,bird’s nest,tower plate,and electric tower,the existing visual inspection system has inconsistent detection accuracy for large and small targets,which performs well on large targets,but usually fail to detect small targets.To this end,the research on multi-size target detection algorithm for transmission line based on deep convolution networks is of great significance to improve the accuracy and efficiency of transmission line inspectionSpecifically,the main contributions of this paper come into three parts（1）Firstly,an ROI pooling method based on bicubic interpolation align is proposed（BiaROIP）.For the basic convolution network corresponding to the ROI region,the bicubic interpolation method is used to obtain the sub-pixel precision convolution feature map,and then a fixed number of grids are used to generate fixed-size features,which can solve the problem of feature position deviation caused by ROI pooling and improve the feature description precision of targets.Using the VGG16 as the backbone,BiaROIP can achieve a mAP of 62.2%on the inspection image test set of transmission line,which is 9.4%higher than ROIP under the same experimental conditions（2）Secondly,a context cascaded region proposal network is present（CCRPN）.Using the VGG16 as the backbone,a cascaded region proposal network is constructed according to the context between the small target and background.The first-stage RPN detects and locates a large-scale background area on the feature map outputted by the Conv5₃,and the second one performs small targets on the convolutional feature map corresponding to the large-scale background area detected by the first-stage RPN,which can enhance the representation ability of feature for small targets.The CCRPN algorithm can achieve a mAP of 66.1%on the inspection image of transmission line test set,which is 5.3%higher than Faster R-CNN under the same experimental conditions.（3）Thirdly,a feature fusion method based on multi-level convolution feature concatenated is proposed（MCFCF）.According to the network structure characteristics of FPN and Faster R-CNN,this paper designs a multi-size target detection algorithm for transmission line based on FPN-MCFCF and FR-MCFCF,respectively,which can improve the adaptability of the detection model to the scale changes of targets.The FPN-MCFCF algorithm concatenates the Top-down feature of the first layer called M5 by bilinear interpolation and the Bottom-up feature of the second layer called C4 to construct multi-scale convolution features,and then input the features to RPN for target detection.The FR-MCFCF algorithm constructs multi-scale convolution features by concatenating the Conv4 and Conv5 convolutional features of VGG16,and then input the features to RPN for target detection.The mAP of the FPN-MCFCF algorithm and FR-MCFCF algorithm have reached 75.1%and 65.7%on the inspection image of transmission line test set respectively,which is 0.8%and 2.8%higher than FPN and Faster R-CNN under the same experimental conditions.This paper conducts research on multi-size target detection algorithm and multi-size target detection experiment for transmission line from the following three aspects including ROI pooling convolutional network features,context region proposal networks between target and background,and multi-level convolution feature concatenate-based fusion Method.The experiments show that the proposed method can improve the accuracy of multi-size target detection and can meet the requirements of multi-size target detection for transmission line.