Small target bullet hole recognition in complex environment based on modified YOLOv5

DOI:	10.11809/bqzbgcxb2024.01.039
Keywords:	deep learning; instance segmentation; YOLOv5; bullet hole recognition; complex environment
Abstract:	In order to recognize bullet holes on target surface efficiently and accurately in complex environments such as different illumination, shadow occlusion and overlap of bullet holes, an instance segmentation algorithm based on improved YOLOv5 was proposed. Based on the YOLOv5 backbone network, the model integrated the segmentation decoder with Atrous Spatial Pyramid Pooling module to realize the instance segmentation of small bullet holes. The algorithm added decoupled detect head to reduce the coupling effect of regression parameters and category probability and improved the recognition accuracy. In addition, the output scale of the mode was adjusted by deleting the large scale forecasting feature layer and adding the minimum scale prediction layer that fused low level information to improve the detection recall rate and accuracy of bullet holes. For the model trained by the data set, the detection accuracy under the test set reaches 92.42%, and the detection speed reaches 25.15 frames/s. Compared with the original network YOLOv5, Mask RCNN, Deeplabv3+ and other networks, the proposed model has higher detection accuracy and detection speedin complex environments and meets the requirements of accuracy and real time in actual training.
Published:	2024-01-28
PDF