Research On Object Detection Technology For Small Objects

The current general object detection algorithm has demonstrated its capabilities in many application scenarios,but there are still great difficulties in object detection in small object scenarios(such as UAV detection,aerospace missions,etc.).The reason is: first,some objects in the image are too small in relative or absolute size,and are easily ignored during object detection.In addition to the size problem,small objects often have problems such as blur,occlusion and aggregation,which presents a great challenge for the neural network to extract the features of small object objects.Therefore,this also makes the results of the current general object detection network directly applied to the small object scene unsatisfactory.In order to solve the above problems,this paper conducts a detailed study on this.The main content are as follows:(1)The YOLOv5 algorithm currently has excellent performance in terms of accuracy,speed,and model size.Now,on the basis of YOLOv5,a small object detection algorithm that improves YOLOv5 is proposed.One is to improve the multi-scale prediction structure of YOLOv5 for the problem that the relative size or absolute size of the small object is too small,and combine 4x downsampled feature maps with YOLOv5’s Neck.Modify the original feature pyramid structure,and finally detect smaller objects through the detection box corresponding to the feature map,so as to improve the detection rate of small objects.The second is to integrate the attention module of the fusion channel and space from the YOLOv5 algorithm.The feature map first uses the channel attention module to adjust the weights between channels to enhance the features of small objects,and then uses the spatial attention to enhance focus on small object areas and reduce the impression of background areas.Finally,through experiments,it can be found that the detection accuracy of the improved method is improved compared with the original algorithm.(2)Aiming at the problems that the relative size of high-resolution image objects is too small in small object scenes and the small object images are generally blurred and lack sufficient appearance information,this paper first introduces the method of combining the image segmentation method with the general object detection network.This paper first introduces the image slice method,and provides a method of combining this method with a general object detection network.Then,a small object detection algorithm based on image super-resolution(SR)is proposed.By introducing the SR-multitask network,the small-size prediction frame detected by the general object detection network as the background is re-detected,and finally the detection results of the SR-multitask network and the detection results of the general network are merged.The small object detection algorithm based on image enhancement combines the above two methods.First,the relative size of the object is enhanced by the image slice method,and then its absolute size and appearance information are increased by the SR-multitask network.Through the final experiment,it can be found that the method based on image enhancement can effectively improve the accuracy of object detection.