Mbouembe Philippe Lyonel Touko, Liu Guoxu, Sikati Jordane, Kim Suk Chan, Kim Jae Ho
Department of Electronics Engineering, Pusan National University, Busan, Republic of Korea.
School of Computer Engineering, Weifang University, Weifang, China.
Front Plant Sci. 2023 Apr 3;14:1150958. doi: 10.3389/fpls.2023.1150958. eCollection 2023.
Automatic and accurate detection of fruit in greenhouse is challenging due to complicated environment conditions. Leaves or branches occlusion, illumination variation, overlap and cluster between fruits make the fruit detection accuracy to decrease. To address this issue, an accurate and robust fruit-detection algorithm was proposed for tomato detection based on an improved YOLOv4-tiny model. First, an improved backbone network was used to enhance feature extraction and reduce overall computational complexity. To obtain the improved backbone network, the BottleneckCSP modules of the original YOLOv4-tiny backbone were replaced by a Bottleneck module and a reduced version of BottleneckCSP module. Then, a tiny version of CSP-Spatial Pyramid Pooling (CSP-SPP) was attached to the new backbone network to improve the receptive field. Finally, a Content Aware Reassembly of Features (CARAFE) module was used in the neck instead of the traditional up-sampling operator to obtain a better feature map with high resolution. These modifications improved the original YOLOv4-tiny and helped the new model to be more efficient and accurate. The experimental results showed that the precision, recall, score, and the mean average precision (mAP) with Intersection over Union (IoU) of 0.5 to 0.95 were 96.3%, 95%, 95.6%, and 82.8% for the improved YOLOv4-tiny model, respectively. The detection time was 1.9 ms per image. The overall detection performance of the improved YOLOv4-tiny was better than that of state-of-the-art detection methods and met the requirements of tomato detection in real time.
由于温室环境条件复杂，自动且准确地检测温室内的水果具有挑战性。树叶或树枝遮挡、光照变化、水果之间的重叠和簇拥会降低水果检测的准确性。为了解决这个问题，基于改进的YOLOv4-tiny模型，提出了一种用于番茄检测的准确且鲁棒的水果检测算法。首先，使用改进的主干网络来增强特征提取并降低整体计算复杂度。为了获得改进的主干网络，将原始YOLOv4-tiny主干的BottleneckCSP模块替换为一个Bottleneck模块和一个简化版的BottleneckCSP模块。然后，将一个微小版的CSP空间金字塔池化（CSP-SPP）附加到新的主干网络上以扩大感受野。最后，在颈部使用内容感知特征重组（CARAFE）模块代替传统的上采样算子，以获得具有高分辨率的更好特征图。这些修改改进了原始的YOLOv4-tiny，使新模型更高效、准确。实验结果表明，改进后的YOLOv4-tiny模型在交并比（IoU）为0.5至0.95时的精度、召回率、得分和平均精度均值（mAP）分别为96.3%、95%、95.6%和82.8%。检测时间为每张图像1.9毫秒。改进后的YOLOv4-tiny的整体检测性能优于现有最先进的检测方法，满足番茄实时检测的要求。
