Research On Algorithms Of Small Object Detection For Images Of Extremely High Resolution

As one of the most fundermental subjects in computer vision field,object detection has a broad prospect for application.In recent years,lots of algorthms for object detection has emerged,achieving great progresses in terms of both accuracy and speed,which in some degree accelerates the process of artificial intelligence industry landing on real life.The popularity of high-performance equipment has lowered the threshold for high resolution image acquisition.However,up to now,due to limitations in hardware computing capability,even the best object detection algorithms are far from meeting the requirements of actual project when dealing with high-resolution images.Considering the resolution of images to be processed being too high as well as the objects to be detected are relatively small and sparsely distributed in project where images are captured by unmanned aerial vehicle during cruise,the most popular algorithms of object detection at present has been tested,and RFB-Net is chosen among single-stage object detection algorithms as basement to be optimized.Progress in speeding up the process has been made by introducing more lightweight convolutional network as backbone,more sparsely distributed square anchor configuration as well as accelerated non-max suppression algorithm,and to increase the accuracy by introduce prediction module based on multi-branch convolution and context feature fusion,which makes the performance demand basically met for our project.The algorithm implemented achieves a speed of 1.43 frames per second with an input image size of 3000*3000,far more faster than the original RFB-Net algorithm's 0.62 frames per second.The loss of precision is small,and the recall is not significantly reduced,which means the accuracy reduction does not lead to additional detection failure.The task of small object detection for images of extremely high resolution requires super fast algorithm,low missing rate,and has smooth demands for precision.Therefore,the algorithm we implement can meet the requirements of our project.