Research On Recognition Technology Of Safety Helmet Wearing In Low Illumination Environment

In the construction environment,the construction personnel wearing safety helmets can effectively avoid or reduce the injury of safety accidents.The machine vision technology realizes the recognition of the safety helmets worn by the workers on the construction site and avoids the occurrence of safety accidents.In harsh environments,the helmet wearing recognition algorithm based on machine vision technology often fails and misdetects.Aiming at the difficulty in recognition of helmet wearing at night and foggy weather,this paper studies image enhancement technology and deep learning technology,and proposes an effective low-illuminance image enhancement algorithm and a lightweight and high-precision helmet wearing recognition convolutional neural network.To achieve accurate recognition of wearing helmets in low-light environments.The research in this paper includes the following aspects:Aiming at the problem of low definition of helmet images in the night environment,this paper proposes an optimized local consistency fusion illumination enhancement method.Aiming at the problems of distortion and unsmooth edges in the enhanced image processed by the multi-scale fusion illumination enhancement method,this article uses the optimized local consistency method to generate a smooth initial illumination map to avoid distortion of the enhanced image.Secondly,this article chooses to use the illumination sub-image is fused with the corresponding weight to produce a refined light map,and the reflection image and the refined light map are fused to obtain a clear image with enhanced illumination.The improved illumination enhancement method can obtain an image with better enhancement effect,and at the same time,the edges of objects in the image are smoother.Aiming at the blurring of the helmet image in a foggy environment,the defogging method based on optimized contrast and the fast defogging method are used to compare the defogging effects.The former achieves defogging by estimating the atmospheric light value and refined transmittance,and the latter a clear image after defogging is obtained by estimating the atmospheric light and ambient light.According to the advantages and disadvantages of their respective defogging,this paper chooses the fast defogging method as the defogging method for helmet images,and adjusts the parameters to adapt to the real foggy environment.Aiming at the problems of missed detection and false detection in existing helmet wearing recognition algorithms,this paper proposes a lightweight and high-precision helmet wearing recognition convolutional neural network.Taking RFBnet-512 as the reference network,in order to improve the recognition accuracy of helmet wearing,the intensive connection processing is used to strengthen the feature extraction ability of the backbone network,and the EFPN feature fusion structure is proposed to enhance the semantic information of shallow features.In order to speed up the recognition speed of helmet wearing,to reduce the number of network channels and deep feature extraction modules to lighten the network.To enhance the network’s ability to recognize occluded targets,CIo U loss and DIo U-NMS post-processing are introduced,the aspect ratio of the network default frame is modified according to the statistical results of the training data,to speed up the convergence speed of network training.This paper considers the convenience and interaction of the safety helmet wearing recognition network,and uses Py Qt5 to develop the safety helmet wearing recognition interface according to the needs.Through experimental comparison,it is concluded that the lightweight and high-precision convolutional neural network proposed in this paper has a good recognition effect on helmet wearing.Compared with YOLO-v3 and RFBnet-512,the recognition accuracy is increased by 3.4% and 2.8%,respectively.The recognition speed is compared with RFBnet-512,it has increased by 69%,and the model file size has been reduced by 39%,which proves the effectiveness of the work in this paper.In the case of night and fog,a clear image is obtained after processing by improved illumination enhancement and defogging algorithms,and then the hard hat wearing recognition network in this paper is used for recognition.