Research And Application Of Small Object Detection Based On Deep Learning

The object detection methods of deep learning are changing with each passing day.However,there is still great room for improvement in the detection of small objects.As a kind of special object,small object is not obvious because of its small size and low resolution in the image,so it is always one of the most challenging tasks in the field of object detection.With the deep integration of artificial intelligence and traditional industries,small object,as an important object in application scenarios,has a wide range of application value.In this paper,the deep learning small object detection method is studied for the two application requirements of routine small object detection and small object counting of intensive bundled steel bar in aerial photography.In the conventional research of small object detection,an improved small object detection algorithm based on feature fusion model was proposed to solve the problem that FPN algorithm ignored the reuse of lower-layer features,which combined the two-way pyramid model of multi-scale feature deep fusion and the detection method of cascade R-CNN framework with gradient threshold.By shortening the information path from lower-layer features to higher-layer features,the feature details of small objects are fused and enhanced to achieve the balance effect of semantic and spatial features.In order to improve the detection effect of small objects,a cascade R-CNN detection network was constructed by introducing a gradient increasing detection threshold structure.The position regression quality of candidate box was optimized by ladder type to enhance the positioning accuracy.The proposed improved algorithm is tested on a conventional multi-scale object dataset,and the experimental results verify the universality and effectiveness of the improved algorithm.This paper applies improved algorithm based on aerial drones the practice of the small object detection routine tasks,the results show that compared with FPN model,the accuracy of the proposed algorithm is improved by 8.1%,improved the aerial transport of small object detection precision and effect.It provides important reference value for traffic monitoring.For easy ambiguity intensive small object detection,detection of defects,using a fusion of gaussian mixture model precision intensive small object detection method,based on Retina Net object detection framework,the introduction of mass fraction of predict box,and with the help of a gaussian mixture model and EM algorithm combined and screening tests,so as to solve the detection problem of ambiguity in densely populated small object detection,and applied in the intensive banding steel bar counting small object tasks,the experimental results show that the method used in reinforcing small object detection accuracy can be up to 93% on average.The Retina Net AP of the baseline method is improved by 3%In counting task,the method compared with Retina Net,in MAE with 59.4,64.2 in RMSE of ascension,raise the accuracy of the reinforced small object count results,realize the detection of box and the real object matching one by one,solved the difficult problem of intensive reinforced accurate count,improve the efficiency of steel bar counting,has certain engineering application value.