Aiming at the problem that existing detection algorithms are difficult to detect small and dense citrus in natural scenes, a DS-YOLO (Deformable Convolution SimAM YOLO) algorithm for dense citrus detection is proposed. Deformable convolution is introduced to extract partial convolution layers of the network instead of the features in the original YOLOv4. The feature extraction network adaptively extracts the location features that result in missing citrus shape information, such as occlusion and overlap. In the feature fusion module, a new detection scale is added and the SimAM attention mechanism is fused to enhance the models ability to extract small and dense citrus features. The results show that the DS-YOLO algorithm improves the accuracy by 8.75%, recall by 7.9%, and F1 by 5% compared with the original YOLOv4 algorithm. It can detect dense citrus targets of the natural environment more accurately and provide effective technical support for dense fruit yield prediction and harvesting robots.

A binarizationbased Faster R-CNN （Binary Faster R-CNN） region detection neural network model is proposed for the study of citrus leaf disease detection and recognition techniques in natural scenarios. The improved model transforms the original Faster R-CNN fully connected layer neural network into a binary, fully convolutional neural network. The experimental results showed that the average recognition rate of the model was 87.2%, 87.6%, 89.8%, 86.4%, and 86.6% for black spot, ulcer, citrus_greening, scab, and healthy leaves of citrus, respectively, and the overall average recognition rate of the model was 87.5%. The recognition speed of the model was improved by 0.53 s compared with the Faster R-CNN network, and the detection time up to 0.31 s per image. The model size was reduced to 153 MB, and the floatingpoint computational power was 2.58×109, while the model converged quickly, ensuring the effectiveness of detection and robustness of the model. The method has good recognition speed and robustness for citrus leaf disease detection in complex natural environments and is of great research significance for citrus disease prevention.

 The deep neural network model is widely used in the identification of plant diseases and insect pests and has achieved great success. At the same time, the computational complexity and parameter quantities of these networks are also increasing, which will pose a major challenge to the deployment of neural networks, especially on devices with limited hardware resources or realtime applications. Aiming to solve this problem, a lightweight pest identification model with Ranger optimizer is proposed, in which the VGG16 is improved by using the Ghost module and reducing the number of convolution kernels in the convolution layer. Experimental results show that the accuracy of the model on the PlantVillage dataset is 99.37%, and the FLOPs is 88.45 M, which is 71.86% lower than VGG16, and it has a faster convergence rate. In a complex environment, the accuracy of the model is 92.40%, and the time is 50% of VGG16.

 Diseases and insect pests of roses seriously affect the yield and ornamental value. The application of a target detection algorithm in the detection of rose diseases and insect pests is conducive to improve the detection efficiency and plays an important role in the realization of the intelligent cultivation of roses. In view of the influence of complex background on the detection of diseases and insect pests in the actual planting scene, a twostage rose pest detection method TSDDP was proposed in this paper. Firstly, the optimized Inception module was added to improve the ability of feature extraction and fusion of YOLOv3 model. The leaf detection of multileaf images of roses in the natural environment was carried out to remove the shadow in the complex background. Then, Faster RCNN was optimized by Kmeans clustering Anchor box to meet the needs of target detection of rose diseases and insect pests. By comparing the detection effects of YOLOv3, Faster RCNN, and TSDDP in the natural environment, the results showed that the detection accuracy and positioning accuracy of TSDDP were higher than other algorithms, and the final average detection accuracy of pests and diseases reached 82.26%. This can effectively reduce the false detection caused by complex backgrounds and improve the detection and location effect of smallscale diseases and insect pests.

In order to study the image recognition technology of citrus in natural environment and realize the early yield prediction of citrus, an improved D-YOLOV3 algorithm was proposed. In this study, a green citrus image data set was constructed. In order to enhance the diversity of the data set, preprocessing operations were carried out on the collected images, including color balance, brightness transformation, rotation transformation, blur, and noise. To solve the problem that gradient information in deep networks will disappear or over-expand with the deepening of the network, the improved model adopts DenseNet's dense connection mechanism to replace the last three lower sampling layers of feature extraction network Darknet53 in the YOLOV3 network to enhance the propagation of features and realize feature reuse. The D-YOLOV3 model was tested by the self-made data set, and experiments were conducted on the recognition performance of the network before and after the modification, different pretreatment methods, different amounts of fruit, and different amounts of data images on the model. The experimental results show that compared with the traditional YOLOV3 model, the accuracy rate of the improved D-YOLOV3 model is increased by 6.57%, the recall rate is increased by 2.75%, the F ₁ score is increased by 4.41%, the intersection ratio is increased by 6.13%, and the average single test time is 0.28 s. Different preprocessing methods enhance the robustness of the model, among which the fuzzy processing has the greatest influence on the performance of the model and the rotation change has the least influence. In the multi-fruit scene, the improved model has a higher recognition accuracy of 5.53% than before, which proves that the model has advantages in recognizing multi-target fruit in the actual scene. Only 1,250 images are needed to fit the model. The research results show that the D-YOLOV3 model proposed in this paper has high accuracy in recognizing immature green citrus in the natural environment, providing technical support for the early production measurement of citrus.