Research On Lightweight Low-illumination Object Detection Method Based On Attention Mechanism

With the advancement of technology,deep learning has been a great success in the field of computer vision in the era of artificial intelligence and big data.Currently,the task of object detection under normal illumination has flourished,but the task of object detection under low illumination still faces numerous challenges.Due to insufficient illumination,the image imaging quality is not high and is susceptible to interference from background and noise,so the detail information is severely lost to meet the expected requirements.In this paper,the YOLO algorithm is used and combined with attention mechanism as the key technique to study object detection in low illumination,which eventually detects test samples such as multi-scale and small objects under low illumination with reduced training parameters.The following two low-illumination object detection methods are proposed in this paper.(1)The SESA-YOLO(Squeeze Excitation Spatial Attention You Only Look Once)object detection algorithm is proposed to address the situation that existing object detection algorithms are not suitable for detection in low-illumination environments.Firstly,the algorithm proposes a new SADarknet network structure,which can differentiate the intensity of feature extraction according to the illumination intensity of low-light images by means of a deep convolutional neural network combined with an attention mechanism to achieve a selective representation of low-light feature maps.Secondly,a new spatial pyramid network DSPP(Deformable Spatial Pyramid Pooling)is constructed to enable adaptive changes in the location of sampling points according to different convolutional kernels under low illumination conditions,in order to adapt to geometric deformations such as shape and size of objects under low illumination,and to meet the detection of complex-shaped objects;In addition,the top-down multi-level feature fusion module can fuse the shallow and deep information of low-illumination images at different levels to enhance the semantic representation of low-illumination feature maps.Finally,GIoU is chosen as the loss function of the bounding box to optimize the network towards the direction of high overlap between the predicted and real boxes.The experimental results show that the SESA-YOLO model reduces both the missed and false detection rates on the low-illumination dataset,and improves both the location information and category accuracy compared with the classical YOLOv3 algorithm.(2)To resolve the conflict between object detection performance and training complexity,the LYOLOv5(Lightweight You Only Look Once)low-light object detection algorithm is proposed.In the paper,firstly,the convolutional layer in the YOLOv5 network is replaced by a GhostNet network to generate an intrinsic feature map by using a small number of convolutional kernels and a simple linear transformation on top of this to generate a rich feature map to overcome the limitation of a small number of low-lightness feature maps.Secondly,a lightweight attention(EPCA)model is designed to allow the selection of perceptual regions to discriminate between background and noise with slightly improved detection accuracy and fewer training parameters.Finally different paths in the jump connection can be adaptively assigned different weights to local regions,allowing effective feature extraction for both high level features and low level features,which in turn captures a large amount of global contextual information in low illumination,resulting in a better representation of the low illumination edge information feature map.Experiments show that the proposed LYOLOv5 network effectively improves the object detection accuracy in low illumination and also reduces the training parameters for low-illumination object detection,making it superior to other lightweight object detection algorithms.