Research On Anchor-free Deep Network Algorithm For Small Object Detection

Deep learning method has significantly improved the object detection performance.However,due to the down sampling of feature extraction,object details will be considerably lost due to the convolution operation of deep network,resulting in objects with small weak information cannot be accurately detected.Therefore,it is of great research significance to design an effective detection algorithm framework to extract small objects features in images and solve the problem of missing detection or false detection.In addition,the annotation cost of data set is high,especially for small object,which is difficult to be detected and easy to be missed.In order to decrease the dependence of model training on the annotated data set,this thesis designs training strategies through semi-supervised learning and active learning,which realizes small object detection with less data annotation cost.During the testing stage,the spatial scale of the objects is in contradiction with the training stage.Therefore,a data pre-processing strategy based on image blocking is introduced.The image to be processed are divided into sub-image sets,which are used to solve the feature change of objects during different reasoning stages.Moreover,to overcome the problem of insufficient training of small objects,a loss function of training sample balance perception is defined.It balances the learning of various scale object samples,and effectively perceives small objects and enhances the detection capability.Experimental results show that the method proposed can effectively improve the performance of small object detection.To overcome the problem of poor detection performance with small amount of training samples,a semi-supervised training mechanism using pseudo labels is proposed.The pseudo labels are generated by binary bounding box level mask,which compares the threshold of pseudo labels with the predicted value of unlabeled objects,and generated the mask matrix for training network model.In addition,the loss function of pseudo label objects is defined for model training,too.Experimental results show that the proposed semi-supervised mechanism can obtain effective information from unlabeled data,and optimize the network model.Particularly,it can reduce the amount of labeled data and the annotation cost for model training.An active learning method is proposed to select high information samples to reduce data annotation cost.Considering the characteristics of the small bounding box set,two different sampling query strategies is used to design the information evaluation indexes correspondingly.The samples with objects difficult to be predicted or in small scale are found in the unlabeled data pool for accurately labeling.With the increase of the amount of training sample size,the model optimization is able to achieve.The experimental results show that making use of the informative samples the network model can be optimized continuously and the detection accuracy of small objects can be improved.In this thesis,as the issues of the poor detection of small objects,the high cost of sample labeling and the difficulty to in obtaining informative samples are studied.The deep learning algorithm framework of anchor-free network is proposed,and the Visdrone2019 dataset is used for object detection experiments.The experimental results validate the effectiveness of the proposed algorithms.