The accurate detection of young fruits in complex scenes is of great significance for automatic fruit growth monitoring systems. The images obtained in the open orchard contain interference factors including strong illumination, blur and occlusion, and the image quality is low. To improve the detection accuracy of young apples in low-quality images, a novel young apple detection algorithm that fuses the YOLOv4 network model and visual attention mechanism was proposed. The Non-local attention module (NLAM) and Convolutional block attention model (CBAM) were added to the baseline of the YOLOv4 model, and the proposed model was named YOLOv4–NLAM–CBAM. NLAM was used to extract the long-range dependency information from high-level visual features; CBAMs were used to further enhance the perception ability of the region of interest (ROI). To verify the effectiveness of the proposed algorithm, 3 000 young apple images were used for training and testing. The results showed that the detection precision, recall rate, average precision and F₁ score of the YOLOv4–NLAM–CBAM model were 85.8%, 97.3%, 97.2% and 91.2%, respectively, and the average run time was 35.1 ms. For highlight/shadow, blur, severe occlusion and other images in test set, the average precision of the proposed algorithm was 98.0%, 96.2%, 97.0% and 96.9%, respectively. The experimental results showed that this method can achieve high-efficiency detection of low-quality images. The method can provide a certain reference for the research on automatic monitoring of young fruit growth.

复杂场景下幼果的准确检测对于自动水果生长监测系统具有重要意义。在露天果园中获得的图像包含强光照、模糊和遮挡等干扰因素，图像质量较低。为了提高低质量图像中小苹果的检测精度，提出了一种融合YOLOv4网络模型和视觉注意机制的小苹果检测算法。在YOLOv4模型的baseline中加入了Non-local attention module（NLAM）和Convolutional block attention model（CBAM），提出的模型命名为YOLOv4-NLAM-CBAM。NLAM用于从高层视觉特征中提取长程依赖信息；CBAM 用于进一步增强感兴趣区域（ROI）的感知能力。为了验证所提出算法的有效性，使用3 000张年轻苹果图像进行训练和测试。结果表明，检测精度、召回率、平均精度和YOLOv4-NLAM-CBAM模型的F ₁分数分别为85.8%、97.3%、97.2%和91.2%，平均运行时间为35.1 ms。对于测试集中的高光/阴影、模糊、严重遮挡等图像，所提算法的平均精度分别为98.0%、96.2%、97.0%和96.9%。实验结果表明，该方法可以实现对低质量图像的高效检测。该方法可为幼果生长自动监测的研究提供一定的参考。