Research On The Improvement Methods For Small Object Detection Based On RetinaNet

The sample imbalance has plagued the object detection tasks for a long time.The sample imbalance in the object detection tasks includes imbalance of the positive and negative samples,imbalance of the difficult and simple samples and imbalance of small,medium and large objects.At present,for the imbalance of the positives and the negatives or the imbalance of the difficults and the simples,many object detection tasks have done well impact.The proportion of small objects in the total training samples is small,which leads to loss imbalance between the smalls,the mediums and the larges,so that the detection model ignores the training of small objects,and the decreased performance of small object detection weakens the overall performance of the detection model.In addition,factors such as the lack of features,the blurring of details and the complex background further weaken the performance of small object detection.The work of Faster RCNN and Retina Net only touches on the imbalance of the positive and negative samples or the imbalance of the difficult and simple samples,and doesn’t explore ways to improve the performance of small object detection.Based on Focal Loss,this thesis proposes an effective improvement scheme for the small object problems in two application scenarios of the Retina Net detection framework(road object detection,face detection),which improves the overall performance of the detection model.The main work of the thesis is as follows:1.In the road object detection scenario,for the loss imbalance among the small,medium and large objects in object detection tasks,a Retina Net Detection Model based on Adaptive Balancing Loss(ABL)is proposed.This model can increase the proportion of small sample loss in the total training loss,so that the direction of Retina Net training is inclined to small objects,and more attention is paid to the training of small objects.Through experimental comparison,the loss-self-balancing Retina Net-ABL has significantly improved the mean average precision of the KITTI dataset,and the detection performance of each category in KITTI has been improved to varying degrees.2.In view of the problems that tiny faces are easily missed in face detection,some face detection models are large and the inference of some model is slow,a lightweight face detection model(GFace)is proposed.Firstly,taking advantage of the Ghost Moudle and depthwise separable convolution,a Ghost bottleneck-v2 structure is proposed combined with the Lambda Layer,and then a lightweight backbone network(Ghost Net-L)is designed.Then,under the Retina Face architecture,combined with the excellent small-sample fitting ability of DCNv2,a lightweight face detection model GFace is designed.Finally,assisted by ABL,the GFace is made lightweight,near real-time and accurate.