Research On The Method Of Lightweight Small Object Detection In Complex Scenes

Object detection is a crucial problem in the field of computer vision,and small objects have become a topic of particular interest in recent years.Traditional detectors often struggle to effectively detect small objects due to their small pixel area and lack of distinctive features.The presence of occlusion,blur,low lighting,and similar objects in complex scenes further exacerbate the difficulty of small object detection.Additionally,traditional target detection models require high-performance hardware platforms due to their numerous parameters and extensive computational requirements,which makes them unsuitable for scenarios with limited cost,volume,and mobility.As a result,developing lightweight models that can meet accuracy and speed requirements using limited computing resources has become a major research focus.Lightweight small target detection technology has the potential to expand the application fields of target detection,such as small defect detection on mobile terminals,small target detection on drones,and traffic sign detection in the field of automatic driving.This paper addresses a series of scientific issues related to the task of lightweight small target detection in complex scenes.The primary objectives of this research include:（1）Small target detection in complex scenes is a challenging problem,as the features of small targets often get submerged by background noise after multiple convolutions,leading to low detection accuracy.Furthermore,conventional target detection models are unsuitable for embedded devices with limited processor computing power.To address these challenges,this chapter presents an exploratory research that focuses on improving the calculation speed while maintaining 90%accuracy.Specifically,we propose an improved lightweight network architecture（SANet）that incorporates SE attention module and Inception structure,enhancing the feature extraction ability of the network in complex scenes and effectively suppressing background noise.In addition,we improve a new feature fusion module（CAM）that efficiently fuses the semantic and position features of different levels by leveraging the spatial position information of the low-level feature layer,enabling the model to distinguish between similar objects.Finally,we improve a lightweight decoupling detection head that improves the model’s ability to decode small objects by decoupling classification and regression tasks.Experimental evaluations on the public COCO2017 dataset demonstrate that the proposed SA-YOLO model,with a parameter volume of only 1.14M,achieves an average small target detection recall rate（AR_S）of 31.6%.（2）Detecting small targets in complex scenes is a challenging problem that existing models struggle to address effectively.Hence,there is a need to enhance the robustness and generalization ability of the models to different scenes.To this end,we improve a Context Enhancement Module（SCEM）that improves the fusion of local and global features,boosting the model’s robustness to spatial layout and object occlusion.Further,we replace the original C3 module with a Stronger Feature Extraction SC2F module in the neck network to enhance the model’s feature extraction capabilities for small objects.We also employ a new loss function and training strategy in the loss function part,emphasizing the weight ratio of small target loss and reducing the negative impact of unbalanced sample size.Moreover,we compress the model and remove redundant network structures by means of channel pruning,thereby decreasing the model’s computational load and improving its generalization ability.We conducted experiments on our self-built dataset IFHB,and the results show that our CE-YOLO model had a parameter quantity of only 0.883M,with an average precision(AP₅₀)of 91.6%.（3）To validate the practical performance of our proposed method for small target detection,we conducted experiments on detecting small targets such as human bodies in a complex factory scene.Our results indicate that our model（CE-YOLO）meets the performance requirements of security products.