Research And Implementation On Algorithms Of Tiny Object Detection For Images Of High Resolution

High resolution images are more and more common in production and life,and objects in these images are often tiny,such as construction machinery in aerial images of high-altitude UAV,ships in satellite images,vehicles and pedestrians in urban monitor video.Tiny object detection in high-resolution images can be applied to many aspects of production and life,such as threat inspection of UAV optical cable,maritime ship recognition,and tiny defects of industrial components detection.There are two difficulties in detecting tiny objects in high-resolution image:one is that the image is too large to use limited hardware resources to directly train and inference the original image;the other is that the size of objects in the image is so small that while scaling the original image to smaller size,the objects is synchronously reduced to several pixels,which makes detection pretty difficult.Based on the actual project,inspired by the process of the human eye identifying tiny objects in high-resolution image,we propose a mask based algorithm for tiny object detection for high-resolution image.The rough detection thumbnail of the mask network gets candidate areas,which are dynamically mapped to the corresponding areas on the original image,after cropped,fed into the detection network for fine-grained detection,and finally the detection results are fused back to the original image,It is called DZNet(dynamic zoom-in network).We take Centernet as the mask network baseline and Yolo V3 as the detection network baseline.Aiming at the task characteristics of large image and tiny object,the backbone network of the mask network,the head network,the neck network and the loss function of the detection network are optimized,the k-means algorithm based on IOU is improved,the NMS based on distance is designed,and the candidate region extraction method combining the candidate region center and the prior information of the candidate region is designed,.according to the characteristics of the data set in this paper,the data augmentation method of fog noise is added.Finally,the performance of the test dataset in this paper is 44.4 m AP,44.4 m AP_S,51.3m AP_M,91.3%recall and 11.7%mistake when the single 4000*6000 size inference time is106.4ms.The performance is better than other advanced object detectors and split based detection algorithms,and meets the performance and speed requirements of the actual project.