Research And Application Of Object Detection Of Transmission Channel Construction Machinery Based On Improved Faster R-CNN

According to the 2014-2018 Safety Production Accidents（Incidents）Analysis Report of the State Grid Corporation of China,one of the main causes of transmission line tripping is the external force damage to the line by construction machinery such as excavators and tower cranes.Therefore,object detection of transmission line construction machinery is of great significance to maintain the stable operation of the power grid and maintain the normal production and life of human beings.We mainly focus on the problems of unbalanced construction machinery category and low detection accuracy of small targets under transmission lines.Based on the existing image data enhancement technology and object detection technology,we propose an improved Faster R-CNN-based object detection method for transmission channel construction machinery.Through improvements at the data level and algorithm level,the detection performance of the model on the task of detecting the construction machinery of the transmission channel is improved.Towards the problem of unbalanced dataset categories,we propose improved solutions from the perspectives of image sampling and data enhancement:（a）Image sampling.We propose a new image sampling algorithm,called the image sampling algorithm based on the TF-IDF weight calculation method.Drawing on the idea of TFIDF weight calculation,this method takes into account both the overall distribution of categories in the dataset and the local distribution in each image in the calculation of image weight,so that the categories with few samples can be sampled for several times,so as to make the categories of dataset fed into the model for training as balanced as possible;（b）Data enhancement.We propose a fine-grained data enhancement algorithm,called the dataset sample expansion algorithm for class balance perception.This method breaks the strong co-occurrence relationship between categories,and can dynamically adjust the number distribution of samples of each category by means of data enhancement during the training process,so as to achieve the effect of balancing training samples.Towards the problem of low detection accuracy of small objects,we first analyze the size distribution of the object in the dataset,and then calculate the distribution of the prior anchor box that is more in line with the object size distribution in the dataset through the K-Means clustering algorithm,and guide the generation of the RPN network a better candidate area.In addition,we also propose a new feature fusion method called fully associative feature fusion with learnable weights.This method can better retain the information of the underlying feature map and add learnable weights to make the model distinguish the importance of different feature maps during feature fusion,and further improve the information fusion capability of the feature pyramid,thereby improving the accuracy of small object detection.In this paper,the Faster R-CNN object detection model is improved at the data level and algorithm level for the problem of category imbalance and low detection accuracy of small objects.The final detection mAP of the model on the test set is 68.1%,which is 8.8%higher than the model without improvement.The model has been applied in real scenes,and the performance of the model is good.