Object Detection Based On Differentiable Network Architecture Search

With the advent of the age of intelligence,object detection technology has been implemented in many fields.For different detection tasks people often rely on networks designed by expert experience,which requires a high level of design and practical skills,and in effect raises the bar for network framework design.Therefore,the use of Neural Architecture Search algorithms to achieve automatic network search has become a hot topic of research today,through the neural network search network has the advantages of small number of parameters,low model complexity and fast detection.The first research component of this paper is the selection of an algorithmic framework for object detection that is most suitable for integration with MobileNet lightweight networks,where the algorithmic frameworks include SSD,YOLOv3,YOLOv4,RetinaNet,and Faster-RCNN.The results on the VOC2007+2012(VOC)dataset show that MobileNet combined with YOLOv4 has the highest m AP to meet the requirements of lightweighting,and therefore YOLOv4 is chosen as the research framework.Although the m AP of MobileNet-YOLOv4 is somewhat reduced compared to YOLOv4,the number of parameters and complexity are reduced by 40%and 20.3% respectively,and the FPS reaches 60.The insulator dataset is subsequently trained in MobileNet-YOLOv4,and the test results show that the model identifies insulators with 97.78% m AP and the average detection speed is faster than YOLOv4 by 0.168 s,and the weight parameter was reduced by 20.91% compared to YOLOv4.The second research component of this paper is to investigate the DifferentNetwork Architecture Search algorithm(DARTS),selecting VOC and COCO2017 as the training data for the study.The DARTS algorithm is used to perform a network search on the above two datasets to search for Reduction Cell and Normal Cell respectively.By replacing the YOLOv4 backbone extraction network with a stacked network of 14 Cells,the number of model parameters and complexity were reduced to some extent.The training method of knowledge distillation is then introduced,using the ResNet101 network as the teacher’s network and the network searched by DRATS on the above two datasets as the student’s network,to improve the model accuracy by distillation.The DD-YOLO algorithm achieves a m AP of 88.56% on the VOC dataset,with a reduction in the number of parameters and complexity of 97.58% and 61.25%respectively compared to YOLOv4,and an FPS of 58.The m AP on the COCO dataset reached 43.5%,the number of parameters and complexity decreased to 97.5% and61.4% of the original,respectively,and the FPS reached 58.In summary,MobileNet-YOLOv4 and DD-YOLO algorithms are improved in terms of detection speed,number of parameters and complexity,so these two different algorithms should be chosen for different application scenarios.