基于DPAA-YOLO v8的复杂逆光场景下柑橘检测方法

Citrus Detection Method in Complex Backlight Scenarios Based on DPAA-YOLO v8

摘要:

针对自然果园环境中强逆光与果实密集分布、枝叶遮挡等深度耦合导致的目标纹理丢失及多尺度特征混淆问题，本文提出一种面向复杂逆光场景的柑橘检测模型DPAA-YOLO v8。首先，该模型以YOLO v8n为基准在骨干网络设计双路融合交叉部分（Dual-path fusion cross stage partial，DPFCSP）模块，通过双分支并行处理策略分别捕捉上下文依赖与局部细节，显著增强模型在逆光暗区及过曝区域的纹理特征提取能力;其次，在颈部网络引入动态自适应注意力（Dynamic adaptive attention，DAA）机制，通过感知输入特征的语义上下文动态调整卷积核权重与感受野大小，有效聚焦前景果实并抑制逆光产生的高频背景噪声;最后，采用Wise-IoU（WIoU）损失函数优化边界框回归策略，解决由光晕效应及阴影干扰引起的目标定位漂移问题。试验结果表明，DPAA-YOLO v8模型平均精度均值（mAP@0.5）达到96.8%，较YOLO v8n提升3.0个百分点;小目标（APs）与大目标（APl）检测精度分别达到57.2%与90.9%，较YOLO v8n分别提升7.8、5.9个百分点。田间机器人集成试验结果表明，系统在强逆光环境下定位成功率为92.5%，平均推理耗时29.2 ms。本文方法在保证轻量化与实时性的前提下显著提升了复杂光照环境下检测鲁棒性，可为柑橘自动化采收提供可靠的技术支持。

Abstract:

Aiming to address the problems of target texture loss and multi-scale feature confusion caused by the deep coupling of strong backlight, dense fruit distribution, and branch occlusion in natural orchard environments, a citrus detection model named DPAA-YOLO v8 for complex backlight scenarios was proposed.Based on YOLO v8n, a dual-path fusion cross stage partial (DPFCSP) module was designed in the backbone network.By employing a dual-branch parallel processing strategy to capture context dependencies and local details respectively, the module significantly enhanced the texture feature extraction capability in backlit dark areas and overexposed regions.Secondly, a dynamic adaptive attention (DAA) mechanism was introduced into the neck network.This mechanism dynamically adjusted convolution kernel weights and receptive field sizes by perceiving the semantic context of input features, effectively focusing on foreground fruits and suppressing high-frequency background noise caused by backlight.Finally, the Wise-IoU (WIoU) loss function was adopted to optimize the bounding box regression strategy, resolving the target localization drift caused by halo effects and shadow interference.Experimental results showed that the mean average precision (mAP@0.5) of the DPAA-YOLO v8 model reached 96.8%, an improvement of 3.0 percentage points over the original YOLO v8n.The detection precision for small targets (APs) and large targets (APl) reached 57.2% and 90.9%, respectively, representing improvements of 7.8 percentage points and 5.9 percentage points compared with that of the original YOLO v8n.Field robot integration tests indicated that the system's positioning success rate in strong backlight environments was 92.5%, with an average inference time of 29.2 ms.The proposed method significantly improved detection robustness under complex lighting conditions while ensuring lightweight and real-time performance, providing reliable technical support for automated citrus harvesting.