基于改进YOLO11n的野外小目标检测算法OA
Wild small object detection algorithm based on improved YOLO11n
针对野外环境中目标尺度小、纹理信息弱、背景干扰复杂以及边界框定位偏差明显等问题,提出一种基于改进YOLO11n的野外小目标检测算法.首先,引入自适应亚像素上采样(ASU)模块,通过亚像素重排和注意力掩码,增强小目标边缘和纹理细节特征;其次,引入多尺度显著性增强模块(MultiSEM),通过多尺度空洞卷积和双池化通道注意力,增强小目标区域响应并抑制复杂背景干扰;最后,构建显著性指数交并比损失函数SE-CIoU,通过引入中心距离指数惩罚和遮挡感知形状约束,提高小目标检测的定位精度和稳定性.实验结果表明,在自建WDD数据集上,改进算法的精度、召回率和 mAP@0.5分别为 88.8%、91.4%和 94.0%,较 YOLO11n基准网络算法分别提高 0.6%、6.6%和3.3%;在iWildCam2022-Small数据集上进行的泛化实验表明,mAP@0.5达到64.5%,较YOLO11n提高3.7%.所提算法能够在保持轻量化检测框架的基础上,有效增强野外小目标的细节表达、显著性响应和边界框定位能力,提高复杂野外场景下小目标检测精度,并具备一定泛化能力.
To address the problems of small target scale,weak texture information,complex background interference,and obvious bounding-box localization deviation in wild environments,an improved YOLO11n-based algorithm for wild small object detection is proposed.First,an adaptive sub-pixel upsampling(ASU)module is introduced to enhance feature fusion for small objects.Through sub-pixel rearrangement and attention masks,ASU strengthens the representation of edge and texture details,thereby reducing the loss of fine-grained information during upsampling.Second,a multi-scale saliency enhancement module(MultiSEM)is introduced before the detection head.By using multi-scale dilated convolutions,the module extracts spatial features under different receptive fields,and dual-pooling channel attention is further used to enhance the response of small target regions while suppressing interference from complex backgrounds.Finally,a saliency-exponential Complete Intersection over Union loss function(SE-CIoU)is constructed for bounding-box regression.By introducing a center-distance exponential penalty and occlusion-aware shape constraints,SE-CIoU strengthens the position and shape constraints between predicted boxes and ground-truth boxes,thus improving the localization accuracy and stability of small object detection.Experimental results show that,on the self-built WDD dataset,the improved algorithm achieves a precision of 88.8%,a recall of 91.4%,and an mAP@0.5 of 94.0%,which are 0.6,6.6,and 3.3 percentage points higher than those of the YOLO11n baseline network algorithm,respectively.Generalization experiments on the iWildCam2022-Small dataset show that the mAP@0.5 reaches 64.5%,which is 3.7 percentage points higher than that of YOLO11n.The results indicate that the proposed algorithm can effectively enhance detail representation,saliency response,and bounding-box localization ability for wild small objects while maintaining a lightweight detection framework.Therefore,the improved YOLO11n algorithm improves small object detection accuracy in complex wild scenes and shows certain generalization ability.
王岑;王彩霞;刘鹏
长春理工大学 电子信息工程学院,吉林 长春 130022长春理工大学 电子信息工程学院,吉林 长春 130022长春理工大学 电子信息工程学院,吉林 长春 130022
信息技术与安全科学
野外小目标目标检测YOLO11n亚像素上采样边界框回归
wild small objectobject detectionYOLO11nsubpixel upsamplingbounding box regression
《液晶与显示》 2026 (8)
1144-1157,14
吉林省科技发展计划(No.20250102228JC)Supported by Jilin Provincial Science and Technology Development Plan(No.20250102228JC)
评论