Research On Target Detection Around Oil Pipelines Based On Computer Vision

In order to ensure the safety of oil pipelines,unmanned aerial vehicles are often used to assist manual inspections of oil pipelines.The inspection flight altitude is high,resulting in a large number of small targets after imaging,and complex scenes in the field may also have targets blocked.At this stage,conventional target detection algorithms have poor detection capabilities for small targets and target occlusions,making it difficult to meet the inspection tasks of field oil pipelines.In view of the requirements of speed and accuracy in the field oil pipeline inspection task,YOLOv4 is selected from the deep learning target detection algorithm as the basic algorithm of this paper,and the detection ability of small targets and occluded targets is improved by improving the YOLOv4 algorithm.The main research content of this paper is as follows:1.Small targets have fewer available features,and their resolution is lower than that of large/medium targets.Feature information is easily lost in the neural network feature extraction stage,which easily leads to low detection accuracy of small targets.Based on this,this paper adds an attention mechanism to the feature extraction stage of the YOLOv4 model to form a YOLOv4-C network,and uses the attention mechanism to enhance the network’s feature extraction and retention capabilities for small targets.The attention mechanism is used in the layer before the feature extraction network finally outputs the multi-scale feature map,so that more detailed information of small targets can be input into the feature fusion network,and the small target information in the feature fusion network can be enriched,thereby improving the accuracy of small targets.Test your ability.The experimental results show that the improved YOLOv4-C method has a good detection effect on small targets around the oil pipeline.2.The surrounding environment of the field oil pipeline is complex,and the target is easily blocked.The detection of occluded targets requires rich semantic information and detailed information.The YOLOv4-C network can extract sufficient detailed information.Considering that there will always be other information interference around the occluded target,a larger detection field of view is required.Based on this,the feature fusion network of YOLOv4 is studied,using the ASPP structure with a larger field of view without increasing the amount of calculation,and improving the field of view by setting the hole rate of the hole convolution reasonably,and the problem of insufficient fusion ability of semantic information and detailed information,replace the PAN structure in YOLOv4 with the RFP structure,use a continuous feedback structure to improve the fusion ability of feature information,and assign weights to each fused feature in the RFP structure,and use learnable weights to improve semantic information and detailed information The fineness of the fusion makes the characteristics of the fusion more focused.The improved and complete YOLOv4-CR algorithm also has a good detection effect on the occluded targets around the oil pipeline.