Research On Small Target Detection And Tracking Based On Improved Deep Residual Network

With the continuous development of imaging technology,the detection,recognition and tracking of objects in the image can effectively help the machine understand the content in the image and lay the foundation for subsequent further analysis.When the target is far away from the optoelectronic device,the imaging size is small.Usually,a target with a length and width less than 10%of the original image size or a size less than 32 × 32 pixels is defined as a small target.The limited resolution and information content of small targets make the research of small targets a huge challenge in the field of computer vision.Small target detection and tracking technology is widely used in many fields such as intelligent interaction and security monitoring,such as the detection of land and sea targets in air-to-ground remote sensing satellite images,and the real-time tracking and warning of all kinds of small flying targets,including unmanned aerial vehicles,are inseparable from the support of image processing technology.Timely monitoring of various potential suspicious targets can ensure adequate response time,which plays a vital role in safeguarding national maritime security,air and space security.The target detection and tracking algorithm based on deep learning technology can greatly alleviate the limitation of over-reliance on prior knowledge and data structure,and has the advantages of intelligence and high automation,which is suitable for processing massive image and video data.On the basis of improving the deep residual network and combining the needs of practical application scenarios,two small target detection and tracking algorithms are proposed in this paper,and the main contents are as follows:First,the detection of small ship targets in air-to-ground remote sensing images.The optical remote sensing image is affected by natural environment and imaging system,and the small ship targets in the complex background are missed and misdetected.In this paper,an improved deep residual network is proposed to extract robust features from active learning of images.We build a dense connection between the shallow and the deep layer of the extracted multi-level pyramid to make up for the lack of spatial location information in deep feature map.A bidirectional feature enhancement network structure combining top-down and bottom-up paths is designed.The shallow details are fully integrated into the feature graph with high level semantic information.At the same time,multi-scale prediction is carried out on the enhanced feature maps at all levels,so that all levels share the parameters of classification and regression subnet,and the classification and regression tasks are completed.Through the verification on the ship dataset,the proposed algorithm can effectively alleviate the problem of missing detection of small targets.Second,tracking low and slow small targets in ground-to-air video sequences.Small flying targets have the characteristics of small imaging size,strong maneuverability,complex and changeable low altitude environment,and difficult to be detected.These characteristics make the detection and tracking of low,slow and small targets in optical images or video become a typical problem of weak target detection and tracking in complex environment.In this paper,a deep network model suitable for video detection and tracking of small,low,slow and small targets is studied,and the network model can perform well under the conditions of small and weak targets,fast target movement and complex background.A re-detection module based on the improved deep residual network is designed to generate fused multi-scale and more discriminative feature maps.The reference frame is used to supervise the process of global search for small targets.On this basis,combined with the tracking module based on depth features,the model is updated regularly,and then a complete set of long-term stable tracking framework for small targets is realized.Experimental results show that the proposed method can achieve better tracking performance.