Research On Automatic Generation And Domain Adaptation Of Small Object Data Set

With the creation of large-scale datasets,data-driven deep learning methods have greatly improved the performance of object detection models.However,in the field of small object detection,due to the difficulty of data acquisition and labeling,such methods are not suitable for small object detection task.This paper uses virtual simulation technology to generate a large-scale complex scene aerial vehicle small object dataset Vehicle Small Object Dataset(VSOD),and develops automatic labeling tools to quickly annotate.However,the existence of domain shift between datasets greatly limits the development of synthetic dataset.This paper designs a domain adaptation model for object detection tasks,which effectively reduces the domain shift between VSOD and real small object datasets.The main research content of this paper is as follows:On the basis of virtual simulation technology,a synthetic vehicle small object dataset construction scheme is proposed.This scheme uses the virtual engine UE4 to build basic virtual scenes of cities,suburbs,and viaducts.Place 12 types of vehicle 3D models and a series of background 3D models in the scene according to the actual application scenario.We use self-designed post-processing tools and 3D rendering methods to obtain synthetic dataset with multiple weather,lighting and shooting angles and use the designed automatic labeling algorithm for accurate 2D bounding box labeling.We studied the domain adaptation method between different domain data,and proposed a pixel-level domain adaptation model for object detection tasks based on the Cycle GAN model.We first improve the Cycle GAN generator,add U-Net structure to the original generator to improve the quality of the generated image.At the same time,the structure of upsampling and convolution is used instead of the deconvolution layer to eliminate the checkerboard effect.After that,a single-stage object detection model of the feature pyramid structure is built to access the improved Cycle GAN model and feature consistency loss is introduced to add effective feature consistency supervision to the data optimization process.The domain adaptation model optimizes VSOD,narrows the domain shift with real dataset,and improves the generalization ability of detectors pretrained with VSOD in real scenarios.We have studied the theoretical basis of fine-tuning of deep neural network,and proposed a multi-stage fine-tuning training strategy.The fine-tuning process is divided into three stages: VSOD pre-training stage,domain adaptation VSOD fine-tuning stage,and target domain dataset fine-tuning stage.This strategy improves the effect of Finetuning in small object detection tasks,and realizes the application of VSOD dataset and domain adaptation model in Fine-tuning.