Study On The Object Detection Algorithm Based On Lightweight Network

Deep neural network-based object identification technologies are widely employed in modern society in domains such as autonomous driving,mobile entertainment,and video monitoring.However,deep learning-based object detection systems are frequently complicated in structure and require a significant amount of computation,making real-time target detection problematic for embedded platforms or edge mobile devices with low computational capabilities.The lightweight network design tries to reduce the number of model parameters and their complexity while retaining model accuracy.This study proposes a lightweight real-time target detection method suitable for embedded platforms or edge mobile devices,as well as a lightweight improvement and optimization design for the YOLO family of networks.To begin,the original YOLOv3 three feature scale detection has been extended to five feature scale detection for small item identification scenarios such as drones,making full advantage of the comprehensive knowledge of numerous scale characteristics to help increase the model’s accuracy.To expedite feature extraction,use the lightweight networks Ghost Net and Sandglass to establish a lightweight network in the backbone network.The best accuracy observed on 1080 Ti can achieve98.92% precision,and the real-time speed is 62.37 FPS during training and validation on the created drone dataset in the dynamic urban scenario.Second,two improved lightweight real-time object detection algorithms,AlphaSGANet and Alpha-EGBNet,are suggested for natural scenes,for datasets such as PASCAL VOC and MS COCO,utilizing YOLOv5 s as the benchmark model.The following enhancements are included in Alpha-SGANet: To begin,add a prediction detection head to detect targets of various sizes,i.e.,add a layer of downsampling(P6)to help increase the receptive field and make thorough use of multi-scale information,keeping cost in mind.Shuffle Net V2 is utilized in the backbone network to create a lightweight and efficient feature extraction network,which uses SPP with smaller convolution kernel sizes to reduce information loss.Then,in the neck section,GAFPN is smartly employed to aid the transition of feature processing,which primarily contains the Ghost module to build effective feature maps to aid prediction,as well as the CBAM module to locate the region of interest in the scene;Finally,the model supervised training achieved the greatest performance gain when paired with Alpha-Io U loss.The experimental results demonstrate that our suggested AlphaSGANet has a real-time speed of 68.49 FPS and an accuracy of 65.14% m AP on the PASCAL VOC dataset,an increase of 7.52% m AP over the original YOLOv5 s.The accuracy on the MS COCO dataset is 39.73% m AP,and the real-time speed is 74.6frames per second.A compact model version,Alpha-SGAs Net,is offered to better balance speed and accuracy in object detection.It achieves 62.62% m AP on the PASCAL VOC dataset with 2.84 MB of parameters,and has a real-time speed of105.3 FPS and 3.01 MB.On the MS COCO dataset,the parameter volume reaches37.84% m AP at a real-time pace of 104.2 frames per second.Alpha-SGANet has a real-time speed of 19.46 FPS on the NVIDIA Jetson AGX platform,while AlphaSGAs Net has a real-time speed of 24.23 FPS,which is 4.92 FPS faster than YOLOv5s’ 19.31 FPS.On the embedded platform,this outcome can meet the majority of the scene’s real-time requirements.Further attempts to refine and optimize Alpha-SGANet are performed,and the Alpha-EGBNet algorithm is proposed.It mostly consists of the following enhancements: To build efficient and rich feature maps while speeding up the network,first use ESNet,an upgraded version of Shuffle Net V2,as the backbone network,and then use SPP to help boost the receptive field in the last-level feature map;The feature transition processing in the neck is done with GA-Bi FPN,which is different from Alpha-way SGANet’s of using four feature scales for prediction.The idea of Bi FPN is combined on the three feature scales to help the fusion of multi-scale feature information,and the Ghost and CBAM modules are also used to effectively improve the model’s performance;finally,combined with Alpha-Io U loss to help supervised model training get more accurate bounding regression boxes.Alpha-EGBNet outperforms Alpha-SGANet on both COCO and VOC data and has real-time detection speed,according to experimental results.