Research On Detection Method Of Chemical Safety Signs Based On Scale Invariance

There is a high danger of accidents in chemical manufacturing plants,especially in chemical manufacturing plants with inadequate safety management.Chemical safety signs are used to remind people of unsafety factors and prevent accidents.They have specific colors and graphical symbols and provide important safety information.They are divided into four types: prohibition signs,warning signs,instruction signs and prompt signs.In order to improve the safety of such plants,this thesis carries out research on the multi-scale detection of chemical safety signs based on the in-depth analysis of relevant methods in the field of computer vision.The goal is to develop a system that can detect the safety signs near workers and automatically notify them.This model is an integral part of small automatic equipment,which can automatically notify and warn workers of the signs in their vicinity.Workers will be informed of potential hazards and can be equipped with necessary protective equipment or avoid dangerous behaviors,which will improve safety and reduce accidents in chemical plants and other fields.How to accurately identify chemical safety signs in real time is the key to realize such an automation device.Because multi-scale sign detection will be the most complex part of such devices,solving this problem basically makes it possible to build such small devices.Because they can be based on hardware similar to cheap mobile phones,they are not too expensive,that is,they are commercially feasible.The most studied field in sign detection is probably traffic sign detection,which has a long tradition and many recent applications.However,there is little research in the field of chemical safety sign detection.Unlike traffic signs,safety signs can be randomly distributed in the field of vision.When many signs,fuzzy signs,small signs or signs with complex backgrounds need to be detected at the same time,the existing sign detection model are insufficient.This thesis develops a new Faster R-CNN variant,which is based on the joint training of two backbones,one for small-scale and one for all signs,which significantly improves the accuracy of small-scale chemical safety sign detection.Although our model and YOLOv5 have achieved about 98% of the total m AP,our model has better small target detection accuracy.Our joint training method provides a solution for the multi-scale detection of chemical safety signs,lays a solid foundation for improving the field of small object detection in the future,and takes another step in the field of sign detection for improving the safety of workers.It can also be applied to other applications,such as remote sensing image detection.We have provided a new benchmark data set to properly train,calibrate and benchmark the detection model.Our research fills an important gap in the research and lays a foundation for the future solid work in the field of sign detection to improve the safety of workers.The main research contents of this thesis are as follows:(1)For each target detection field,the role of standard reference data set is very important.This thesis publicly provides the China Chemical Safety Signs(CCSS)data set,which contains 4650 photos of chemical safety signs from 30 different categories.These photos were taken at different angles,distances,lighting conditions and different scenes.Each sign has a class label and boundary box with annotations and other image features.We have created such benchmark data and publish it together with all source codes of a baseline of trained convolutional neural network(CNN)models and all results in an immutable online archive.Public repository with all data on zenodo.org.Information removed for double-blind review.This thesis has carried out routine test and analysis on the data set,and proved that the data set can be used to evaluate detection algorithms of chemical safety signs,which is a significant contribution to the research community.(2)We comprehensively analyze and compare the performance of the most advanced deep learning models on the data set,including YOLOv3＿spp,SSD,YOLOv5,Faster R-CNN(M,R,V),and Faster R-CNN variant.(3)We have developed an improved version of Faster R-CNN,which is particularly suitable for the detection of small-scale signs using the joint training of two network branches.The average detection accuracy of the improved method for small targets is68.04%,which is 7.01% higher than that of the current best method.Therefore,our research marks an important step forward in the field of sign detection and improves the safety of workers.