Green citrus detection in citrus orchards provides reliable support for production management chains, such as fruit thinning, sunburn prevention and yield estimation. In this paper, we proposed a lightweight object detection YOLOv5-CS (Citrus Sort) model to realize object detection and the accurate counting of green citrus in the natural environment. First, we employ image rotation codes to improve the generalization ability of the model. Second, in the backbone, a convolutional layer is replaced by a convolutional block attention module, and a detection layer is embedded to improve the detection accuracy of the little citrus. Third, both the loss function CIoU (Complete Intersection over Union) and cosine annealing algorithm are used to get the better training effect of the model. Finally, our model is migrated and deployed to the AI (Artificial Intelligence) edge system. Furthermore, we apply the scene segmentation method using the “virtual region” to achieve accurate counting of the green citrus, thereby forming an embedded system of green citrus counting by edge computing. The results show that the mAP@.5 of the YOLOv5-CS model for green citrus was 98.23%, and the recall is 97.66%. The inference speed of YOLOv5-CS detecting a picture on the server is 0.017 s, and the inference speed on Nvidia Jetson Xavier NX is 0.037 s. The detection and counting frame rate of the AI edge system-side counting system is 28 FPS, which meets the counting requirements of green citrus.

Citrus fruit detection can provide technical support for fine management and yield determination of citrus orchards. Accurate detection of citrus fruits in mountain orchards is challenging because of leaf occlusion and citrus fruit mutual occlusion of different fruits. This paper presents a citrus detection task that combines UAV data collection, AI embedded device, and target detection algorithm. The system used a small unmanned aerial vehicle equipped with a camera to take full-scale pictures of citrus trees; at the same time, we extended the state-of-the-art model target detection algorithm, added the attention mechanism and adaptive fusion feature method, improved the model’s performance; to facilitate the deployment of the model, we used the pruning method to reduce the amount of model calculation and parameters. The improved target detection algorithm is ported to the edge computing end to detect the data collected by the unmanned aerial vehicle. The experiment was performed on the self-made citrus dataset, the detection accuracy was 93.32%, and the processing speed at the edge computing device was 180 ms/frame. This method is suitable for citrus detection tasks in the mountainous orchard environment, and it can help fruit growers to estimate their yield.

The rapid and accurate identification of citrus leaf diseases is crucial for the sustainable development of the citrus industry. Because citrus leaf disease samples are small, unevenly distributed, and difficult to collect, we redesigned the generator structure of FastGAN and added small batch standard deviations to the discriminator to produce an enhanced model called FastGAN2, which was used for generating citrus disease and nutritional deficiency (zinc and magnesium deficiency) images. The performance of the existing model degrades significantly when the training and test data exhibit large differences in appearance or originate from different regions. To solve this problem, we propose an EfficientNet-B5 network incorporating adaptive angular margin (Arcface) loss with the adversarial weight perturbation mechanism, and we call it EfficientNet-B5-pro. The FastGAN2 network can be trained using only 50 images. The Fréchet Inception Distance (FID) and Kernel Inception Distance (KID) are improved by 31.8% and 59.86%, respectively, compared to the original FastGAN network; 8000 images were generated using the FastGAN2 network (2000 black star disease, 2000 canker disease, 2000 healthy, 2000 deficiency). Only images generated by the FastGAN2 network were used as the training set to train the ten classification networks. Real images, which were not used to train the FastGAN2 network, were used as the test set. The average accuracy rates of the ten classification networks exceeded 93%. The accuracy, precision, recall, and F1 scores achieved by EfficientNet-B5-pro were 97.04%, 97.32%, 96.96%, and 97.09%, respectively, and they were 2.26%, 1.19%, 1.98%, and 1.86% higher than those of EfficientNet-B5, respectively. The classification network model can be successfully trained using only the images generated by FastGAN2, and EfficientNet-B5-pro has good generalization and robustness. The method used in this study can be an effective tool for citrus disease and nutritional deficiency image classification using a small number of samples.

The accurate and rapid acquisition of fruit tree canopy parameters is fundamental for achieving precision operations in orchard robotics, including accurate spraying and precise fertilization. In response to the issue of inaccurate citrus tree canopy segmentation in complex orchard backgrounds, this paper proposes an improved DeepLabv3+ model for fruit tree canopy segmentation, facilitating canopy parameter calculation. The model takes the RGB-D (Red, Green, Blue, Depth) image segmented canopy foreground as input, introducing Dilated Spatial Convolution in Atrous Spatial Pyramid Pooling to reduce computational load and integrating Convolutional Block Attention Module and Coordinate Attention for enhanced edge feature extraction. MobileNetV3-Small is utilized as the backbone network, making the model suitable for embedded platforms. A citrus tree canopy image dataset was collected from two orchards in distinct regions. Data from Orchard A was divided into training, validation, and test set A, while data from Orchard B was designated as test set B, collectively employed for model training and testing. The model achieves a detection speed of 32.69 FPS on Jetson Xavier NX, which is six times faster than the traditional DeepLabv3+. On test set A, the mIoU is 95.62%, and on test set B, the mIoU is 92.29%, showing a 1.12% improvement over the traditional DeepLabv3+. These results demonstrate the outstanding performance of the improved DeepLabv3+ model in segmenting fruit tree canopies under different conditions, thus enabling precise spraying by orchard spraying robots.

Litchi florescence has large flower spikes and volume; reasonable control of the ratio of male to female litchi flowers is the key operational aspect of litchi orchards for preserving quality and increasing production. To achieve the rapid detection of male and female litchi flowers, reduce manual statistical errors, and meet the demand for accurate fertilizer regulation, an intelligent detection method for male and female litchi flowers suitable for deployment to low-power embedded platforms is proposed. The method uses multi-teacher pre-activation feature distillation (MPFD) and chooses the relatively complex YOLOv4 and YOLOv5-l as the teacher models and the relatively simple YOLOv4-Tiny as the student model. By dynamically learning the intermediate feature knowledge of the different teacher models, the student model can improve its detection performance by meeting the embedded platform application requirements such as low power consumption and real-time performance. The main objectives of this study are as follows: optimize the distillation position before the activation function (pre-activation) to reduce the feature distillation loss; use the LogCosh-Squared function as the distillation distance loss function to improve distillation performance; adopt the margin-activation method to improve the features of the teacher model passed to the student model; and propose to adopt the Convolution and Group Normalization (Conv-GN) structure for the feature transformation of the student model to prevent effective information loss. Moreover, the distilled student model is quantified and ported for deployment to a field-programmable gate array (FPGA)-embedded platform to design and implement a fast, intelligent detection system for male and female litchi flowers. The experimental results show that compared with an undistilled student model, the mAP of the student model obtained after MPFD feature distillation is improved by 4.42 to 94.21%; the size of the detection model ported and deployed to the FPGA-embedded platform is 5.91 MB, and the power consumption is only 10 W, which is 73.85% and 94.54% lower than that of the detection models on the server and PC platforms, respectively, and it can better meet the application requirements of rapid detection and accurate statistics of male and female litchi flowers.

Citrus pests and diseases severely impact fruit yield and quality. However, existing object detection models face limitations in complex backgrounds, target occlusion, and small target recognition, and they struggle to be efficiently deployed on resource-constrained devices. To address these issues, this study proposes a lightweight pest and disease detection model, YOLOv11-RDTNet, based on the improved YOLOv11n. This model integrates multi-scale features and attention mechanisms to enhance recognition performance in complex scenarios, while adopting a lightweight design to reduce computational costs and improve deployment adaptability. The model introduces three key enhancement features: First, shallow RFD (SRFD) and deep RFD (DRFD) downsampling modules replace traditional convolution modules, improving image feature extraction accuracy and robustness. Second, the Dynamic Group Shuffle Transformer (DGST) module replaces the original C3k2 module, reducing the model’s parameter count and computational demand, further enhancing efficiency and performance. Lastly, the lightweight Task Align Dynamic Detection Head (TADDH) replaces the original detection head, significantly reducing the parameter count and improving accuracy in small-object detection. After processing the collected images, we obtained 1382 images and constructed a dataset containing five types of citrus pests and diseases: anthracnose, canker, yellow vein disease, coal pollution disease, and leaf miner moth. We applied data augmentation on the dataset and conducted experimental validation. Experimental results showed that the YOLOv11-RDTNet model had a parameter count of 1.54 MB, an mAP50 of 87.0%, and a model size of 3.4 MB. Compared to the original YOLOv11 model, the YOLOv11-RDTNet model reduced the parameter count by 40.3%, improved mAP50 by 4.8%, and reduced the model size from 5.5 MB to 3.4 MB. This model not only improved detection accuracy and reduced computational load but also achieved a balance in performance, size, and speed, making it more suitable for deployment on mobile devices. Additionally, the research findings provided an effective tool for citrus pest and disease detection with small sample sizes, offering valuable insights for citrus pest and disease detection in agricultural practices.