Research On Lightweight Small Object Detection Method Based On Feature Fusion And Attention Mechanism

Small object detection is one of the important research directions in the field of computer vision,which is widely used in scenes such as impurity detection,low-altitude unmanned aerial vehicle warning,industrial product defect detection,and long-distance pedestrian detection,and has high practical value.Due to the limited available feature information and low resolution of small objects in images,small object detection remains a challenging research area.Existing small object detection methods can be categorized into four major types: multi-scale representation,context information,super-resolution,and region proposal.These methods are often complex in structure to achieve high detection performance,which leads to an increase in parameters.However,in limited computing resources environments,many small object detection methods are difficult to deploy and achieve real-time detection.Therefore,it is a valuable research direction to achieve real-time and accurate small object detection while reducing the demand for computing resources.In addition,the lack of specialized datasets for small object detection and the imbalance in sample distribution in existing open-source object detection datasets also limit the development of small object detection research.This thesis addresses the following issues: how to effectively obtain feature information of small objects to improve model detection performance,how to design lightweight modules that focus on target area information,how to optimize the performance of lightweight small object detection methods,and how to enrich small object detection data.The work done in this thesis is as follows:(1)A study of network structure design and feature fusion techniques for small object detectionTo achieve real-time detection of small objects with limited computing resources,this thesis uses the SSD network as the backbone and adds a specialized prediction head for small object detection.Based on research on multi-scale representation and feature fusion,a shallow feature pyramid structure is designed to extract feature information from the backbone network output and fuse it with the semantic information of deep feature maps.The network’s ability to detect small objects is increased while adding minimal computational overhead.(2)Research on the design of context attention modules and loss functionThis thesis proposes a lightweight multi-scale context module for extracting contextual information of small objects,as well as a lightweight attention module that enhances object information using channel and spatial attention mechanisms.By combining with a shallow feature pyramid,an efficient lightweight network for small object detection is achieved,greatly improving the deployability and real-time performance of the network.Additionally,a scale focal loss function is proposed to improve network stability in cases of imbalanced sample distribution in most datasets.(3)Research on knowledge distillation technique based on class score and attention mechanismThe module design for lightweight network ensures real-time and deployability,but limits the network performance.This thesis studies knowledge distillation techniques based on class score and attention mechanisms to guide the student network to learn the feature distribution from a better-performing teacher network,resulting in improved results and enhanced lightweight network performance.The proposed knowledge distillation technique is applied to lightweight small object detection network,achieving performance enhancement.(4)Drone-vs-Bird detection grand challenge and construction of datasetThe low-altitude unmanned aerial vehicle warning task is one of the hotspots in small object detection.This thesis constructs a UAV small object dataset for the low-altitude unmanned aerial vehicle warning task,based on the Drone-vs-Bird Detection Grand Challenge organized by the international conference ICASSP2023,to address the problem of insufficient small object data in small object detection tasks.The dataset contains four different sizes of UAVs and provides manual annotations for these UAVs.This thesis uses the constructed UAVs small object dataset and the proposed small object detection network to achieve a top-five ranking in the Drone-vs-Bird Detection Grand Challenge organized by ICASSP2023.