Research On Small Object Detection Method Based On Feature Enhancement And Contextual Information Aggregation

Object detection technology is one of the important research directions in the field of computer vision,which can be applied to many fields such as autonomous driving,security monitoring,medical diagnosis,smart homes,and aerospace.This research is of great significance for promoting the progress and application of computer vision technology.Due to the small size of small objects,limited pixels,unclear object features,and easy interference from other background noise,small object detection is a difficult point in object detection.In order to improve the performance of small object detection,this paper conducts research on the anchor-based one-stage detection algorithm,and proposes improvements from the following three aspects.(1)A novel auxiliary feature enhancement network is proposed to effectively enhance the representation of small object features.In multi-scale detection networks,shallow feature maps with small receptive fields are commonly used for detecting small objects.However,small objects have small sizes and weaker feature representations,which limits the model’s ability to perceive them based on inherent pixel-level features.To address this issue,this paper presents a novel auxiliary feature enhancement network that leverages the advantages of image pyramids to efficiently extract original image features at multiple resolutions.By incorporating bidirectional fusion operations,including top-down,bottom-up,and lateral connections,into a multi-scale fusion pyramid,the semantic features of small objects are enriched,thereby improving their feature expression capabilities.The effectiveness of the proposed method is validated on three datasets,namely PASCAL VOC,TGRS-HRRSD,and MS COCO.The method achieves a detection speed of 58 frames per second,while significantly improving the small object detection accuracy by 7.0%,2.9%,and 4.9% on the respective datasets.(2)A self-adaptive feature weighting method,focusing on foreground features,is proposed.In images,the target objects often occupy only a small portion,and a large amount of interference information can cause blurriness at the edges of small objects.They are prone to being submerged in the background,making it necessary to enhance the model’s attention to small targets.In light of this,this paper introduces a self-adaptive feature weighting method to differentiate between features and preserve the representation of small targets.This method employs channel attention without dimensionality reduction and spatial attention that focuses on local features in parallel,effectively handling the fine-grained features of small objects and suppressing background interference.Furthermore,it adaptsively adjusts the fusion coefficients based on the contribution of feature maps to the detection results,resulting in prediction feature maps with more effective small target feature information.Ultimately,on the PASCAL VOC and MS COCO datasets,the small object detection accuracy is improved by 7.5% and 6.8% respectively.(3)A context information aggregation method based on the spatial correlation of associated objects is proposed.In natural environments,targets do not exist in isolation but have some form of correlation with their surroundings.The feature information of small targets is limited,and the surrounding pixels can serve as auxiliary information to assist in detecting small targets.Existing methods relying on local context fail to capture the global information of the target object or introduce additional computational complexity.To address this,this paper presents a long-range context information acquisition method based on individual pixels.This method extracts context information along two cross paths at each pixel,mapping it to the original feature map to establish long-range information correlations.The enhancement of regional context effectively improves small object detection.Ultimately,on the PASCAL VOC dataset,an m AP of 81.8% is achieved.Compared to baseline models SSD,Yolov4,Yolov5,and Yolov7,the small object detection accuracy is improved by 8.5%,2.1%,3.7%,and 2.4% respectively.