Research On Small Object Detection For Transmission Lines Based On Convolutional Neural Network

As a key technology of visual inspection for transformation lines,object detection has made a breakthrough.Nowadays,deep learning has become the mainstream method of object detection.However,in visual inspection images of transmission lines,the areas and features of small objects such as hammer,spacer bar,tower plate,and bird's nest are small,resulting in low detection accuracy of Faster RCNN and SSD and other classical object detection architectures.Hence,this paper proposes a research topic on the small object detection algorithm for transmission lines based on the convolutional neural network,which has important practical value for accuracy and efficiency improvement for the visual inspection of transmission lines.The contributions for this thesis mainly come into four parts:First,an adaptive morphological constraint method for small object RoIs is proposed.K-means clustering method is adopted to cluster the size and normalized length and width of the training samples in the transmission line inspection dataset,and then the ranges of the sizes and aspect ratios of the small objects are obtained to estimate the parameters of RPN.Experiments show that the adaptability of RPN to the scale variation is improved.Second,a small object detection method based on context-encoding network is proposed to solve the problem of low feature resolution and poor description accuracy for small objects.In this method,multi-scale context information is aggregated through the dilation convolution module.The feature maps of the middle layer and the top layer are integrated by the deconvolution module.The context is gradually encoded into the subsequent feature maps.Experiments show that our method improves the detection mAP of small objects by 6.70%.Third,a small object detection method based on salient context extractor is proposed to solve the problem of poor feature separability of small object RoIs.This method proposes candidate context at first,and then adds the salient context extraction layer on the top of RoI Pooling layer to obtain the most discriminating context from the candidate context.Then the salient context features and the original RoI features are concatenated for feature fusion.Experiments show that our method improves the detection mAP by 5.33%.Moreover,adding the context-encoding network and the salient context extractor to the pipeline simultaneously can improve the detection mAP by 9.95%.Finally,aiming to enhance the poor representations of small objects,a small object detection method based on super-resolution feature generator is proposed.Generator introduces features from the shallow layer of the feature extractor,learns residual features between small objects and large objects,and integrates with features from the top layer of the feature extractor to construct super-resolution features similar to those of large objects.The discriminator distinguishes the pooled large object features from the generated small object features to guide the generator learning.According to different sizes,transmission line inspection image dataset is divided into large object subset and small object subset.Then the adversarial training is carried out between these two subsets.Experiments show that our method improves the detection mAP by 10.10% on the small object subset and 7.31% on the large object subset.