Small Object Detection Based On Multi-scale Context Information Fusion

Nowadays,computer vision technology is widely used in daily life and engineering field.After years of development,object detection tasks based on deep learning have achieved satisfactory results in large and medium-sized object detection tasks of natural images,but poor performance in small object detection tasks.At the same time,with the progress of remote sensing technology,the innate advantages of remote sensing images in the field of environmental monitoring,national defense and security are becoming more and more obvious,remote sensing images are also applied in more fields,but there are a lot of small objects in remote sensing images.Therefore,many practical tasks put forward higher requirements for small object detection,which is of practical significance to the research of related small object detection technology.At present,the research difficulties of small object detection based on deep learning mainly include:small object features are rare and difficult to extract;Small objects appear in dense distribution,which intensifies the difficulty of object detection.The lack of small object datasets,especially the lack of remote sensing small object datasets,unable to targeted training model.There is a semantic gap between natural images and remote sensing images,and many models pre-trained in natural images cannot achieve the best performance in remote sensing images.In summary,for the current research short board and the actual demand,this paper takes remote sensing small object detection as the main research content and focuses on small objects of 10-20 pixels.The main work is summarized as follows:1.Aiming at the shortage of remote sensing small object dataset,a remote sensing aircraft small object detection dataset is built,and the sample is complex,which has good applicability.All goal instances have been artificially annotated and repeated,providing data support for subsequent research.2.A multi-scale feature adaptive weighted fusion framework was proposed to solve the problem of small object feature extraction:the weighted fusion between different scale features is carried out by network learning.The framework makes full use of the information of each dimension of the feature graph and their contribution to the final fusion feature graph,and then obtains the feature graph with fuller and more perfect information expression,which helps neural network to extract and learn more features of small objects.3.Based on the aforementioned work,a lightweight remote sensing image small object detection method based on multi-scale feature fusion is proposed.Compared to the current mainstream detection algorithm,the amount of model parameters of the method is greatly reduced,and real-time object detection can be achieved.At the same time,small object detection accuracy is also greatly leading other algorithms to meet the actual engineering needs of small object detection of some remote sensing scenes.4.In response to the problem of the dense small object positioning is not accurate and serious problems of missing detection,a multi-scale context information fusion small object detection algorithm based on attention guidance is proposed.By extracting the contextual information assisted network around the object for small objectives,the network focuses on learning the object itself,and reducing the impact of the surrounding similar objects for their own characteristic learning through the attention mechanism.Effectively improve the accuracy of small objects.