Research On Small Object Detection Algorithm Based On Deep Learning

As one of the most important tasks in the field of computer vision,object detection is the basis of other visual tasks.Because deep learning can extract many complex features,it is widely used in target detection.Target detection based on deep learning has been applied in some industrial life,such as driverless,intelligent transportation,medical diagnosis and so on.Although deep learning based target detection is widely used,there are still some problems,such as low detection rate,low accuracy and low recognition rate.In order to solve the problems in small target detection,this paper proposes two improved algorithms based on FCOS(full progressive one stage object detection),which are the improvement of regression strategy and the improvement of feature fusion network,and proposes a new detection algorithm,CFCOS(Cascade FCOS).Firstly,this paper analyzes the advantages and disadvantages of FCOS,and proposes an improved regression method to solve the problem of low detection accuracy of small objects.This method inherits the idea of FCOS algorithm.Instead of using a single grid center point to predict the object,it combines the vertices of the grid to predict the object.In the post-processing process of network prediction,the post-processing process for the prediction results of mesh vertices is correspondingly added to get the final detection frame,and the calculation of the prediction results of mesh vertices is also added in the calculation of loss function.This method greatly reduces the missed detection of small objects and can adapt to more scale objects.Then,in the feature fusion network of FCOS,the imbalance between the abstract information and the detailed information is caused by the over sampling of the top-level feature map and the only top-down fusion path.In this paper,an improved method of multi-scale feature fusion is proposed.A bottom-up fusion path is added on the top-down basis,which helps to balance abstract and detailed information.Then,in order to further fuse the information of different scale feature map,the feature fusion network is cascaded to increase the fusion degree.This feature fusion network can improve the ability of small target classification and location.Finally,this paper uses the method of decomposition and integration through the experiment on mscoco.In the experiment of regression strategy improvement,the average accuracy of the improved method is 1.1% higher than that of the non improved method;In the multi-scale feature fusion experiment,the improved method proposed in this paper has an average accuracy of 1.6% higher than that of the original method on small objects;the combination of the two improved methods is the cFCOS proposed in this paper,which has an average accuracy of 3.8% higher than that of FCOS on small objects.In conclusion,the method proposed in this paper has a good improvement in the effect of small object detection...