Research On Small Object Detection Algorithm Based On Deep Learning

With the development of artificial intelligence technology,small target detection has become a research hotspot in aviation and satellite earth observation missions in recent years.Generally,the object whose boundary box overlaps with the image accounts for 0.08% to 0.58% is called a small target.Small target detection becomes an urgent problem to be solved in civil and military fields in the future due to the lack of effective feature information and insignificant feature.Therefore,in order to improve the accuracy of small target detection,a small target detection algorithm based on deep learning Network and attention mechanism is studied.The main research contents include:In current small target detection algorithms,small target detection is a challenging task in computer vision.Small target detection has low accuracy and poor localization ability due to low resolution and less information,and the loss or reduction of small target features after the pooling layer.Therefore,in order to improve the accuracy of small target detection,a small target detection algorithm based on deep learning Network and attention mechanism is studied.The main research contents include:(1)Aiming at the problem of low resolution and small amount of information contained in small target detection,a strategy of combining Squeeze-and-Excitation Networks(SE)and atrous spatial convolution pooling pyramid is proposed,and UNet is used to capture features.The ability of context information and the characteristics of fast detection speed,detect small objects in a pixel-by-pixel manner,and improve the local feature capture and detection capabilities of small objects.The experimental results show that compared with YOLOv4,m AP is improved by 14% in detecting two types of small objects such as clock mouse.(2)Aiming at the problem that the small and medium target features in the input image to be detected are lost or reduced after the pooling layer,resulting in low detection accuracy and poor localization ability of the small target detection algorithm,a sub-pixel convolution is proposed to enhance the feature map information,and use the atrous convolution block and SE attention mechanism to further improve the contextual feature capture and detection capabilities of small objects.Numerical experiments show that compared with the Retina Net method,the sub-pixel convolutional coding block improves the performance of the small object detection algorithm by1%～2%.(3)The attention mechanism enables the small target detection Network to accurately focus on the context information of the detection target,which has become an important means to improve the detection performance.Although spatial attention and channel attention mechanisms can improve the mean average accuracy of small target detection Network,they have limited ability to capture the context feature information of small objects,which makes the detection accuracy of small targets and large and medium-sized targets gap,and it is difficult to detect the location of small objects.In this paper,an algorithm module based on the channel self-attention mechanism(SCA)is constructed.After the input feature mapping is compressed,the self-attention mechanism is used to establish the correlation between channels and adaptively optimize the response of feature channels,thus improving the ability of small objects to capture the remote context information.Thus,the detection accuracy of small target is improved.The results of numerical experiments show that the m AP values of Res Net based on channel self-attention mechanism and Mobile Net V2 are increased by 1.2% and 1.7% respectively on PASCAL VOC2007 dataset with almost no computational cost increase.It improves the performance of small target detection Network in small object detection.