Study On An Improved YOLOv3 Object Identification Algorithm

As the most advanced target detection algorithm,YOLOv3 can process video data in real time on the most advanced graphics cards and detect targets from it.Therefore,we choose to research and improve on this algorithm.This study consists of two parts,one part is to obtain new anchor boxes through more advanced clustering algorithm under the premise of understanding the operation mode of YOLOv3.The other part is to update and improve on the original basic network,and try to obtain higher accuracy through retraining to prove that the improvement of anchor boxes is effective.In the research and improvement algorithm part,YOLOv3 draws on the improvement of the anchor boxes mechanism used by the PRN network and BBox regression.Its function is to generate object proposals in the network and generate Bounding Boxes through filtering.This part directly affects the accuracy of Bounding Boxes.And it affects the level of mAP-50,which affects the performance level of the whole algorithm.However,based on the results of traditional clustering algorithms,it is often difficult to obtain a stable optimal solution.Therefore,an improved clustering algorithm is used to Performing statistical clustering again is a direction worthy of research.The third chapter of this paper details the ideas of improvement and the results of improved statistics.In the test part,it is also divided into two parts for testing and checking.One part is to update the anchor boxes and fine-tune the network when comparing the author's training environment as much as possible,and compare the updated version with the original test results.The second part is to make the private dataset with annotations,modify the YOLOv3 configuration file,and then retrain the network with the private dataset.After repeated parameter adjustment training,we get a stable minimum loss weight result,and finally use the weighted file matching the network after training.Check its predictive ability,and compare the prediction effect with the network matching the original weight,thereby checking whether the update of the anchor boxes is valid on other data sets,the generalization ability,representative ability,etc.under actual circumstances,this experiment It can directly reflect the influence of anchor boxes on the YOLOv3 target detection algorithm,which is of great significance for improving YOLOv3.