Research And Application Of Oilfield Tiny Object Detection Technology Based On Deep Learning

With the help of computer vision and the rise of intelligent video surveillance system has been successfully applied to all walks of life,by positioning in the image and identify the category of the objects so as to achieve intelligent judgment.It is difficult to detect tiny objects in complex environment in actual oilfield operation site.Therefore,in view of such problems,this paper researches on how to detect tiny objects in complex oilfield field environment,and proposes a deep learning-based detection model for oilfield tiny objects.The main tasks are as follows:(1)The oilfield datasets are constructed on the collected images,and the image defogging algorithm is used for preprocessing to solve the problems of blurring and low contrast caused by fog in the image,so as to improve the image sharpness.By analyzing the principle of the current mainstream defogging algorithm,the atmospheric scattering model is established based on the dark channel defogging algorithm based on image restoration,and the fog-free image is restored.The image defogging preprocessing can provide high-quality images for the subsequent detection model of tiny objects and help the object to be easily located and detected in the image.(2)An improved YOLOv3-MFA(YOLOv3-Multiscale Fusion Attention)algorithm was proposed based on the integration of CBAM(Convolutional Block Attention Mechanism)and the addition of multi-scale feature fusion method.By comparing the current mainstream object detection algorithms,the YOLOv3 object detection algorithm,which has excellent performance in both accuracy and efficiency,is selected as the basic model.However,due to the small inherent pixel ratio of the object in the research scene,the YOLOv3 algorithm still needs to be improved.By introducing the attention mechanism CBAM,the network can capture and locate the object needing attention more attentively,and enhance the information expression of the tiny object.On this basis,according to the idea that the small receptive field in the shallow network is beneficial to the object positioning,and the large receptive field in the deep network is beneficial to the extraction of the semantic information of the object,the multi-scale fusion is carried out by adding scales to make full use of the shallow information,so as to realize the feature reuse.At the same time,the K-means++ algorithm is improved for clustering to find the anchor frame which is more consistent with the characteristics of the self-built datasets,so as to accelerate the convergence speed of the model.The experimental results show that YOLOv3-MFA algorithm improves the detection effect of tiny objects.(3)Design the oilfield intelligent video monitoring system.Based on the above improved algorithm YOLOv3-MFA,the oilfield video monitoring system is designed to conduct real-time monitoring of tiny object such as cigarette end and mobile phone existing in the oilfield operation site,so as to the production area smoking illegal use of mobile phones and other dangerous behavior to take the initiative to warn,reduce the consumption of human and material resources,and improve the efficiency of oilfield safety management.