Design And Application Of Lightweight Object Detection Models

Object detection technology is one of the important research directions in computer vision.Among them,lightweight object detection algorithms suitable for mobile and embedded devices have become a hot topic of concern.However,traditional object detection techniques are unable to achieve fast and efficient detection on such resource-constrained hardware devices.Therefore,improving the efficiency of object detection algorithms is crucial in addressing this problem.This thesis is based on deep learning methods and is dedicated to researching how to improve the algorithmic efficiency of object detection models while achieving a good balance between detection speed and accuracy.The main contributions and innovative achievements of this thesis include the following points.(1)This thesis presents YOLOX-Lite,a lightweight object detection network based on improved YOLOX.Building upon the YOLOX baseline network,several key enhancements are introduced.Firstly,a lightweight attention module called Mixed Efficient Channel Attention is designed to adaptively refine features,emphasizing valuable information while suppressing irrelevant distractions.It possesses the characteristics of plug-and-play integration and minimal additional computational overhead.Secondly,Mobile Netv3 is optimized to obtain a highly efficient backbone network known as Modified-Mobile Netv3.Lastly,an efficient downsampler called Efficient Down＿Sampler with Focus is devised to effectively fuse low-dimensional feature layers from the backbone network with highdimensional feature layers from the neck,achieving a feature layer that encompasses both high and low-dimensional information.When constructing the feature enhancement network using PAnet,depthwise separable convolutions are applied to reduce computational complexity.Experimental results demonstrate that YOLOX-Lite,as implemented in this study,significantly improves the inference speed of the model on the PASCAL VOC2007 and TT100 K datasets,albeit with a slight loss in m AP.(2)This thesis presents Rapider-YOLOX,a high-accuracy lightweight object detection network that improves the detection precision of lightweight networks.Based on YOLOXNano as the baseline network,this study explores methods to enhance the accuracy of lightweight network models.Firstly,a Highly Efficient Bottleneck module is designed to enhance the feature extraction capability of the original YOLOX-Nano model’s depthwise convolution module.Secondly,a Soft-SPP module is devised to prevent the phenomenon of partial important information loss that can occur in the original SPP module.It further improves the network’s ability to fuse multi-scale information and facilitate communication between channels.Lastly,the CIOU(Center Intersection over Union)is introduced as a new loss function to enhance the positional accuracy of predicted boxes by considering the center distance and aspect ratio between predicted and ground-truth boxes.Experimental results on the PASCAL VOC2007 dataset demonstrate that the proposed Rapider-YOLOX model achieves a significant improvement in m AP compared to the original model.It can also achieve real-time detection on a GT1030 with only 384 CUDA cores,with almost no impact on FPS performance.This proves that the proposed method effectively enhances the detection accuracy of the network while maintaining its lightweight characteristics.(3)This thesis presents a object detection system for demolition sites and designs different detection modes based on the characteristics of different detection tasks,including image,video,and real-time camera-based detection.The system allows users to adjust the Io U threshold and confidence values according to specific requirements.Additionally,the system provides real-time display of detection results,enabling users to promptly assess the detection status.Through the application of object detection in demolition sites,the proposed method has been validated and achieved good detection performance,providing valuable references for the practical application of object detection in real-world scenarios.