| With the rapid development of Internet technology,More than 85% of information is in the form of pixels on the Internet.Computer vision technology plays an increasingly important role in human life,and the research on computer vision is also more and more popular.Object detection,as the cornerstone of the field of computer vision,is increasingly used in real life applications,such as target tracking,video surveillance,information security,and automatic driving,Image retrieval,medical image analysis,network data mining,drone navigation,remote sensing image analysis,defense systems,and so on.At present,the commonly used image target recognition algorithms in the market needs artificial image preprocessing and feature extraction on the sample images.The processing results of the algorithms are easily affected by the noise with background,light,and size of the image.This paper uses the Object detection algorithms of deep learning as the main research content,and applies the improved Faster R-CNN algorithm to the monitoring video image of the construction sites.It is used to detect whether the workers wear helmets or not,which helps to Strengthen supervision of site safety production in construction sites.The main content of this article is as follows:1.Described the design process of the convolutional neural network(CNN)model in detail.From the modeling of a single neuron to the design of conventional networks,and finally to the construction of CNN,the origin model and working principle of the CNN which is suitable for image feature extraction are gradually elucidated.2.Introduced the principle of object detection and related details processing in the detection process,explain the current popular object detection depth network algorithms.And analyzed the three methods based on candidate regions from R-CNN to Faster R-CNN,two regression-based methods YOLO and SSD.Finally,used the VOC public dataset to test the Faster R-CNN,Analyzed and summarized the experimental results of the algorithm’s average accuracy,image processing efficiency,and image processing effects.3.Improved the training processes of the algorithm in the experiment,and added hard and negative sample mining strategies in the training set to improve the detection accuracy ofthe experimental algorithm.Training,testing and evaluating the model by the experimental samples based on improved algorithms.Test YOLO,SSD,Faster RCNN,and improved algorithm using experimental test sets,evaluate the performance of the improved algorithm according to the results of each algorithm on the test samples.And the improved algorithm was successfully applied to the project helmet detection in construction sites.The following are the innovations of this paper:(1)In the project of helmet detection in construction sites,deep neural network algorithm is used to replace the traditional image feature processing and classification methods,simplified the complicated picture preprocessing process and improving the generalization ability of the system.(2)Improved the training process of Faster R-CNN algorithm.According to the experimental data set,in the training process,the training strategy of hard negative mining is designed to improve the detection accuracy of the algorithm.(3)The detection model which is applied to the helmet project can identify both the helmet and the worker at the same time,thereby judging the relationship between the detection targets. |
