Research On Detection Methods Of Small Objects In UAV Images Based On Deep Learning

In general,object detection is to find the specified object in the image,and at the same time,it is necessary to find out the object’s position and size.The fundamental content of object detection is ”where is the object”.There are generally two definitions for small objects.One is defined in the coco data set,where the pixel value of the object is less than32×32,and the other is that the width and height of the object are both less than one-tenth of the original image,satisfying the above two One of these conditions is defined as a small goal.With the development of science and technology,the availability of high-resolution aerial images is getting higher and higher,which contains a large amount of information related to safety,land development,disease control,defect localization,surveillance,and so on.However,these data are highly unstructured,so it is difficult to extract useful infor-mation from large-scale data,and even requires intensive manual analysis.For example,the classification of urban land requires professionals to operate.Therefore,this task must be inefficient and expensive.Most of the objects in UAV images are small objects.In or-der to apply deep learning to small object detection in UAV images,this thesis chooses the open source data set Vis Drone as the training set.Aiming at the difficulties in UAV image data collection,this thesis proposes an im-proved method.This thesis mainly works from the following aspects:(1)Aiming at the inaccurate positioning of small objects in the object detection model,this thesis designs an attention mechanism.The self-attention mechanism in this thesis is very flexible and can be flexibly integrated into other detection models.And verify the effectiveness of the attention mechanism in this thesis in the one-stage and two-stage methods.(2)This thesis proposes a specific data enhancement operation(specific)to solve the problem that the light and dark changes in the UAV image make the detection of small objects difficult.(3)Aiming at the problem that the artificial setting of the Anchor size in the object detection model often doesn’t match the small object in the dataset,this thesis proposes the use of Anchor adaptive generation method to make the generated Anchor size more match the current data set to improve the detection of small objects Accuracy.(4)In response to the common sample imbalance and category imbalance problems in the real world,this article proposes a method to adjust the loss function,using focal loss to improve the sample imbalance problem in the dataset.(5)Aiming at the common phenomenon of overlapping detection frames in the ob-ject detection data set,this thesis proposes to use the Softer-NMS strategy to solve this problem.This thesis has done a full experiment on the proposed module on the Vis Drone dataset,and verified the help of the various improvement schemes proposed in this thesis for the difficulty of small object detection in UAV images.