基于感受野增强与多模块协同优化的目标检测算法OA

In object detection tasks,the YOLO series algorithms have achieved an outstanding balance between speed and accuracy.However,there remains room for improvement in their receptive field design.Conventional approaches typically enlarge the receptive field by increasing convolutional kernel sizes,which substantially raises computational complexity.Meanwhile,traditional self-attention mechanisms often rely on local or global windows to compute attention weights,making them susceptible to interference from irrelevant regions and consequently degrading detection accuracy.To address these limitations,this paper proposes YOLO with receptive field enhancement(RFE-YOLO),a novel object detec-tion algorithm based on receptive field enhancement and multi-module collaborative optimization.RFE-YOLO introduces a C3K2_WT(cross stage partial with kernel size 2 and wavelet transform convolution)module in the backbone network,which employs wavelet transformation to achieve an expanded receptive field without excessive parameterization.C2DAT(cross stage partial with deformable self-attention transformer)module that utilizes deformable self-attention to focus on relevant regions and capture more critical feature information is designed.The neck network incorporates a selective boundary aggregation(SBA)module to fuse low-level boundary details with high-level semantic features,enabling pre-cise boundary preservation and accurate localization recalibration.A Down module is introduced to perform downsam-pling operations,effectively reducing model complexity.Experimental results on the benchmark MSCOCO 2017 dataset demonstrate that RFE-YOLO consistently outperforms other YOLO variants,with mAP improvements ranging from 1.1%to 11.6%.Moreover,its mAP is on par with that of the state-of-the-art RT-DETR(real-time detection transformer)family,while the parameter count of RFE-YOLO is only 7.1%of RT-DETR-L.Ablation studies further validate the effectiveness of each proposed module.