Research On Image Object Detection Of Field Patrol System Based On Deep Learning

With the emergence of image and video big data and the improvement of the computing power of GPU-like components,the development of the theory and technology of deep learning has been promoted,and the development of computer vision has reached a new height.Image target detection technology is also widely used in many fields of artificial intelligence.Among them,the research on image target detection methods and techniques has achieved remarkable results,and has become a research hotspot of computer vision,and has achieved great improvement in performance.The recognition and understanding of massive image data can be applied to many aspects such as security,medical care and transportation,and has wide application value.Image target detection in the field inspection system has a series of features,such as a wide range of target angles,high image resolution,different target size,weather image quality,image target blur,target shape and image background change,etc.Based on the deep learning convolutional neural network framework,the paper improves the image target detection algorithm and optimizes it to solve the difficulties in the inspection of target images.The main work and results are as follows:(1)Comparative study on the application of several main depth learning algorithms in surveying line images.Combined with the characteristics of the photographs of the field inspection system,the RCNN,Faster R CNN,YOLO and SSD model algorithms were compared.The experimental process and results show that these algorithms are superior to traditional image detection methods in the field detection of field inspection images.The performance of Faster R CNN and YOLO is more prominent.(2)For the characteristics of candidate frame in Faster R CNN algorithm,a multi-feature and multi-objective improved target detection model(MMF R CNN)algorithm based on Faster R CNN is proposed.The model algorithm is improved,the feature is integrated and the RPN is focused on,the target feature weight is increased,and the background interference in the line image is reduced.The IRPN and the patrol image target detection network are trained at the same time,and the target detection model of MMF R CNN is obtained.Experiments show that the detection accuracy mAP of the MMF R CNN model on the improved data set KITTI has been greatly improved,but the detection time has not increased.(3)In the line image,since the imaging device is long from the target distance,a small target detection problem occurs.The improved E-YOLO algorithm based on regression YOLO target detection algorithm is a target detection algorithm based on full convolution network.The algorithm improves the problem of inaccurate detection of small targets and poor positioning effect in YOLO algorithm.Because it is not affected by input scale,the detection flexibility of YOLO algorithm is increased.At the same time,a multi-border prediction method is improved to improve the detection speed FPS of the small target.(4)The GAN-based data set extension solves the problem of small sample data sets.Deep learning usually requires a large amount of data as a support,and the amount of data is small,which often causes problems such as over-fitting.The public data set is large,but there is no data set that is completely suitable for this topic.In this research experiment,based on the published KITTI data set,combined with more than 1000 real-time survey line images,the data set expansion technology is used to solve the small data set problem.To further verify that the proposed detection algorithm has good generalization capabilities,a series of related experiments were performed in the improved KITTI data set.Through the comparison and analysis of the experimental results,the designed and improved convolutional neural network based inspection image target detection method can achieve good results in both accuracy and speed.