Research On Lightweight Object Detection Technology Based On Feature Optimization

With the rapid development of mobile internet,mobile devices,and multimedia technology,images,as an important information carrier,are increasingly being applied to people’s lives and work.Target detection technology,as one of the important fields of computer vision,can automatically detect objects in images or videos and has become an important technical means in fields such as unmanned aerial photography,smart homes,smart security,autonomous driving,agriculture,and industry.With the continuous development of artificial intelligence and the increasing demand for application,target detection technology will have a wider range of applications in the future,bringing more convenience and benefits to people’s lives and work.In recent years,target detection technology has received much attention from academia and industry,and there are two main types of methods:methods based on handcrafted features and methods based on deep neural networks.Methods based on handcrafted features have problems with poor generalization and robustness,high computational complexity,etc.,while methods based on deep neural networks can automatically learn high-level feature representations and semantic information,thus having better generalization and robustness,and reducing computational complexity and time cost.The current research trend is to further improve the detection performance and efficiency of lightweight models through deep learning techniques and optimization strategies,while considering the portability and adaptability of lightweight models,so as to achieve better application effects in different scenarios.The main directions include the design of lightweight network structures,feature fusion optimization,neural network architecture search,etc.In order to improve the running efficiency and detection performance of lightweight target detection models,this paper focuses on feature fusion and model lightweight research,and proposes a method based on singlescale feature fusion optimization to solve the problems of structural complexity and high computational complexity of traditional multi-scale feature fusion methods.This method uses a dense dilated encoder,which not only enhances the feature extraction ability of large and medium-sized target features but also maximally preserves the information of small-sized targets.It only uses the single-scale features output by the main feature extraction network for feature optimization and strengthens the feature transmission in the atrous convolution module through dense skip-layer connections,thus improving the detection performance of the model.Secondly,a feature fusion optimization method based on lightweight fusion attention is proposed.By optimizing the main network,detection head,and dense dilated encoder modules,the parameter and computational complexity of the model are reduced.At the same time,a lightweight attention weighting mechanism is introduced into the feature optimization module,and adaptive convolution kernel size is used to better capture the importance information between different channels in response to the changing number of input features,which improves the effect of feature fusion.These methods have important theoretical significance and practical application value for solving the problems of high computational complexity,insufficient accuracy,and large model parameters in target detection technology.The experimental results show that on the MS COCO dataset,compared with the one-stage target detection model RetinaNet,DDE-Net achieves a 0.2-point lead in AP indicators while having smaller parameters and GFLOPs than RetinaNet.Compared with the YOLOv5n,YOLOXNano,YOLOv5s and YOLOX-tiny models,LDDE-Net weighs the relationship between the amount of model parameters,the amount of calculation and the detection effect,while ensuring a better detection effect,it reduces the parameters as much as possible.The use of large quantities improves the efficiency of model operation and reasoning.LDDE-Net has achieved an index of 32.5AP while the amount of parameters is only 3.6M,and has achieved a detection score of 15.7 FPS on the mobile embedded platform Jetson Nano.