Object Detection Method In Small Data Set Scenarios

In recent years,object detection technology has been widely used in various fields to fulfill various visual detection tasks.Currently,most of the relatively mainstream target detection algorithms are based on deep learning,which requires a large amount of data to support the training of better performance models.However,for some special target detection application scenarios,such as military security,medical diseases,security emergencies,and other scenarios,obtaining a large amount of data is relatively difficult,How to use small data sets to train robust detection models is very important and meaningful for these special scenarios.Therefore,this article focuses on improving the accuracy of target detection in small dataset scenarios.The main work and innovations are as follows:(1)Aiming at the problems of sparse samples and single types in small datasets that cannot meet the full training of deep learning target detection algorithms,resulting in low detection accuracy,a target detection data enhancement method for small datasets was designed from the perspective of data source enhancement.By combining single image data enhancement and multiple image data enhancement,and using the proposed Mosaic-7 data enhancement method,the number and diversity of data sets are expanded and enriched.Subsequently,the data sets expanded by the two methods are jointly input into the training set to meet the requirements of improving the learning ability of local and global features and improving the model detection effect.And through testing on the captured dataset,the comparative analysis of the results shows that this method can effectively improve the average accuracy(2)Aiming at the problems of false detection and missed detection in the YOLOv7 algorithm in small dataset target detection,a YOLOv7-Swin T target detection method for small dataset target detection was designed from the perspective of improving the detection network.By disassembling and optimizing the YOLOv7 target detection algorithm,the replacement positioning loss function is α＿SIo U,splicing the Swin Transformer classification network,and other operations successively complete the positioning target task and classification target task,and then fuse to complete the target detection task.Through experimental comparison,it can be verified that the modified algorithm can effectively improve the error detection,missed detection,and other situations.(3)From an application perspective,an emergency safety target detection project is presented.Introduced the project requirements,framework,use of Tensor Rt to accelerate model testing,and demonstrated the testing interface.Based on the data sets in the project,the performance and effect of the above improved algorithm under the premise of training in small data sets and YOLOv7 under large data sets are compared.Figure [47] Table [8] Reference [92]...