Research On Weighted Multi-Scale Object Detection Based On Deep Learning

Object detection,as a hotspot of theory and application in computer vision in recent years,is widely used in many fields such as unmanned driving,intelligent video surveillance,aerospace and so on.With the great success of deep learning in image classification,object detection technology based on deep learning has developed rapidly in the past decade,a series of excellent detection models such as R-CNN and YOLO are created.Affected by various factors in the real environment,how to apply real-time object detection to the real environment still has huge challenges:(1)The application scenarios of object detection are often highly time-sensitive;(2)The detection accuracy of the model has high requirements in the real environment;(3)The number of objects in the real environment is large and complex,and the small objects in the image account for a large proportion,the features of small objects are difficult to extract;(4)There are many interference factors in the real environment: rain,fog and other weather objective factors will cause noise to the image.In view of the multiple problems in the above,this thesis implemented an end-to-end target detection model YOLO?FM(YOLO Flexible Module)based on the YOLOv3,the main research contents are as follows:(1)Using the YOLOv3 model as the basic framework,maintaining the advantage of processing image speed;(2)YOLOv3 uses the mean square error as the loss function of the bounding box regression,and IoU is used as the evaluation standard in the evaluation of positive and negative samples,so there is a gap.Therefore,this thesis uses a more reasonable GIoU as the loss function of the bounding box to improve the accuracy of object detection;(3)For small targets in the image,YOLO?FM model uses multi-scale weighted feature fusion to extract the features of small targets to satisfy the needs of small target detection accuracy;(4)According to the impact of the real environment on the detected image,the defogging algorithm is used for preprocessing before the image enters the network to reduce the noise in the image and make the feature of the target in the image more obvious.The YOLO?FM model is comprehensively verified on the VOC2012 and COCO data sets,and the model is compared and analyzed from multiple aspects.The model reached an inference speed of 35 fps on the VO2012 dataset and 33 fps on the COCO dataset,with real-time object detection capability;The improved GIoU-based YOLOv3 model improved 2.2% on mAP relative to YOLOv3,and the GIoU-based YOLO?FM model improved 4.1% on mAP relative to YOLOv3;The model uses multi-scale weighted feature fusion to be more sensitive to small target detection than networks that do not use weighted feature fusion;After using the defogging algorithm,the model has improved accuracy by 5.4% and the fps has decreased,but it still has the ability to detect in real time.