Research On Small Object Detection Algorithm Based On Attention Mechanism

With the improvement of video surveillance systems,the popularity of drone applications,and the development of 5G technologies,the amount of image data collected by devices has grown rapidly,and the requirements for data processing have also gradually increased,which challenges the object detection algorithms.The small size of the objects reduces the detection accuracy,but improving the detection accuracy increases the amount of calculation,which directly limits the application scenario of the detection algorithm.At the present,most of the object detection algorithms deployed in various applications perform static inference.When detecting objects of differcnt sizes,all pixels in the image will be processed.The algorithm spends equal processing time on each pixel of different scales,making the entire process inefficient.Using image pyramids or adding anchors to the RPN network make this problem even worse.In recent years,the algorithms proposed on the small object detection problem have focused on the direction of the attention mechanism,dividing the target detection into two phases.The first phase,named the region generation phase,implements the attention mechanism and the second phase performs the region-byregion object detection.By ignoring the background,algorithm focuses on small objects and only up-samples the area around these objects.But the common problem is the method of implementing the attention mechanism is too complicated.The overall speed of the model has not been improved or even dropped significantly,which means the full advantage of the attention mechanism has not been taken.The small object detection algorithm proposed in this paper differs from other small object detection algorithms in the first phase of detection.We further decompose the first phase into the region perceptual task and the sub-region generation task.Based on the semantic segmentation network,we design region perceptual model,aiming to quickly obtain the region of interests through a simple model.This method reduces the number of processed pixels,and greatly reduces the difficulty of small objects detection.The purpose of keeping high accuracy and speeding up detection is achieved.In the process of sub-region generation,for the scene with a large number of small objects,image morphological operation and the clustering algorithm are used to deal with the unreasonable area size,and to control the number and the size of areas within a proper range.The experiments were carried out on two datasets,TT100 K and Vis Drone.The experiments show that the recall and the accuracy of the first stage of our small object detection algorithm is significantly higher than that of object detection models.Focusing on areas where objects may exist significantly reduces the requirements for the model during the object detection phase.High accuracy and high detection speed is achieved at the same time.On the TT100 K dataset,our algorithm can reduce the number of processed pixels in the model by at least 80% comparing with the models proposed in the recent years,and the average processing time of each image is only 12 ms,which means at least ten times speed increase can be obtained.The size of the region prepceptual model is smaller 10 MB.Finally,28.59% of AP is reached on the Vis Drone dataset,which also indicates the high performance of our algorithm in scenarios where the density of objects is high.