Research On Small Object Detection Algorithm Based On Improved YOLOv5 Model

Small object detection has important and extensive theoretical research worth and practical application value.Due to the particularity of the small target,the detection has always been unsatisfactory.And many object detection algorithms have high requirements for hardware equipment such as the storage capacity and computing power,and cannot be deployed on mobile devices like drones.Therefore,considering the actual application scenario,while improving the detection performance of the model for small object,the computation and calculation of the model can not be too large.Considering the detection performance and speed simultaneously,this thesis has se-lected YOLOv5s as the benchmark model which has more applications on actual engi-neering projects.The specific improvements are as follows:1.In the view of few available features of small objects,this thesis first strengthens the utilization of information resources of small object.The specific improvement points are divided into two directions.First of all,The global context module in backbone can bring more abundant context information of small object for the feature maps of each con-volutional layer,which can to some extent compensate for the problem of limited feature information for small object,so as to provide more clues for the detection of small ob-ject;Secondly,the convolutional attention is added to the neck to enhance small object’s features from both channel and spatial dimensions,which enable the model’s attention to focus on small object regions and enhance the model’s anti-interference ability.The two improvements can achieve efficient utilization of information resources.2.In response to the problem of difficult extraction of small object’s features,this the-sis enhances the model’s feature fusion ability,and the specific improvements are mainly from two aspects.Firstly,the nearest neighbor interpolation during the upsampling pro-cess is replaced by the transposed convolution module,which can also achieve upsam-pling effect and continuously update and adjust the convolution kernel weights by the backpropagation process of the neural network,so as to enhance the fitting ability of the neural network.In order to utilize shallow feature maps with richer positional informa-tion,we can deepen the upsampling multiples in the feature pyramid structure,and obtain an another prediction head that can be used to detect small targets with smaller resolutions.The multi-scale detection range of the model is also expanded,and the problem of missed detection of small object in the model is effectively improved.At the same time,the orig-inal three prediction layers of the model contain more positional information,effectively improving the overall detection performance of the model.This thesis has verified that the relevant improvements are effective in different degrees through experiments,and when all improvements are added at the same time,the model achieves best.Compared with the benchmark model,the AP@.5 on the Vis-Drone2019 testing datasets increased by 4.9%,the AP increased by 3.8%,the AR in-creased by 8%,AP_Sincreased by 1.33%,AR_Sincreased by 7.1%.The parameters and flops of the model separately increased by only 1.74M and 7.5G.The detection speed is50.7FPS,which meets the real-time detection standard.