Research On Low-Light Object Detection Method Based On Improved YOLOv5

As a foundational task in the field of computer vision,the introduction of deep learning techniques has made Convolutional Neural Network representative of mainstream object detection algorithms,with extensive applications in areas such as security inspection and intelligent transportation.However,in practical detection scenarios,the quality of low-light images collected by devices is often poor due to insufficient lighting.Problems such as uneven lighting,insufficient brightness,and unclear details in dim environments can cause the common object detection algorithms to fail to achieve the detection accuracy under normal lighting conditions,posing significant challenges.Therefore,this paper proposes two low-light object detection methods based on improved YOLOv5 to improve the poor detection performance in low-light environments.(1)The first method proposed in this paper is an improved object detection method based on attention and adaptive feature fusion.To optimize the enhancement algorithm,a mixed enhancement training strategy is employed by investigating the impact of low-light enhancement algorithms on downstream object detection tasks.By mixing the original dataset with the dataset enhanced using Enlighten GAN,the model’s generalization ability is improved,and the low-light enhancement algorithm is better adapted to downstream object detection networks.In terms of network structure improvement,by introducing an improved coordinate attention mechanism in the backbone,the network can improve the integration capability for global information while accurately capturing target location information.On the other hand,a composite receptive field feature enhancement module was designed in the neck of the network,and an adaptive feature fusion module was introduced before the output head to enhance the detection effect of occluded objects and multi-scale objects.Experimental results show that the improved method helps to improve the network’s detection performance in low-light environments.(2)The second method proposed in this paper is a lightweight object detection method based on Mobile Vi T and lightweight improvements.Based on the mixed enhancement training strategy,YOLOv5 is improved using Mobile Vi T,a hybrid architecture combining Convolutional Neural Network and Transformer,and two lightweight methods for further network structure optimization.Firstly,a depth residual convolution module based on the fast Fourier transform is proposed to extend the perceptual field of shallow networks at a low cost.Secondly,Ghost convolution modules are introduced at the neck of the network to reduce network structure complexity,and the decoupled fully connected attention mechanism is employed to integrate global information at the feature fusion stage.Experimental results show that the proposed method not only effectively improves detection performance in low-light environments by addressing weak global information capturing ability but also facilitates network structure lightweight improvement.