Improvement Of YOLOv5 Algorithm And Its Practical Application

Object Detection is one of the most basic tasks in Computer Vision,and it is also a prerequisite for completing other complex tasks related to computer vision.Finding and locating a specified object in the information contained in the picture is its main task To complete this task,it is necessary to integrate advanced technologies in many fields,such as deep learning,feature extraction,image processing,and pattern recognition.The challenge is self-evident.More and more target detection algorithms are appearing in research and application fields,such as public safety,medical image processing,unmanned driving,etc.This is not only inseparable from the rapid development of deep learning related technologies,but also closely related to the increasing computing power of GPUs.Related.The current mainstream target detection algorithms include SSD,R-CNN,Fast R-CNN,Faster R-CNN and YOLO algorithms.But from the perspective of detection speed and detection accuracy,just like its name You Only Look Once,YOLO is a coexistence of speed and accuracy.The development of the YOLO algorithm to YOLOv5 can basically be called a fairly powerful algorithm.But it still has shortcomings for the detection of small targets.This article has made an improvement on its original Mosaic enhancement method.The original four pictures are used for random rotation,cropping,zooming,etc.,and then spliced to nine pictures,which is called Mosaic-9,which increases the size of the image.The amount of target data,on the other hand,enriches the background of the detected object and reduces the number of batch?normalization calculations.In addition,on the basis of Mosaic-9 data enhancement,the training technique Label Smoothing is adopted,which effectively solves the problem of label over-confident.Using the improved YOLOv5 in the diabetic foot detection project,m AP reached 97.4%,which is an increase of more than 5 percentage points compared with the previous study.In recent research applications,deployment on mobile devices or front-end embedded devices has gradually become a hot spot.Taking our diabetic foot diagnosis research as an opportunity,we also take the deployment of mobile devices and development boards as one of our research contents.Among the four models of YOLOv5,YOLOv5 s,YOLOv5m,YOLOv5 l,and YOLOv5 x,YOLOv5s is used in the deployment of diabetic foot diagnosis system due to its simplicity and light weight.On the one hand,we converted the trained diabetic foot detection model,and finally packaged it into an apk and installed it on the Android platform;on the other hand,we also completed the deployment of the RK3399 Pro development board.The accuracy and speed can basically meet the actual application.Claim.