The experiment is conducted on rubber trees tapped for one, three, and five year(s). In the National Natural Rubber Forest in Danzhou City, Hainan Province, China, 2029 photos are collected using image acquisition equipment, a Sony Alpha 6000 camera with a resolution of 4000×6000. In order to ensure the richness and diversity of the samples, multi-angle shooting methods are used under different lighting conditions, and photos of 9 scenes are collected. It includes rubber trees with one, three, and five year(s) of tapping age. The rubber trees of each tapping age also include rubber trees that block the end point of tapping but not the starting point of tapping, rubber trees that block the starting point of tapping but not the end point of tapping, and rubber trees that both the starting point and end point are blocked at the same time, as shown in Figure 1 . Finally, Labelme image annotation software is used to manually label the rubber tree’s tapping area, tapping starting point, and tapping endpoint to create a JSON format data set.

In deep learning network model training, the richness, diversity, and accuracy of the data set have a decisive impact on the final training results of the network model. The singleness and deficiency of the data set will lead to the model being overfitted. At the same time, due to the complex environment of the rubber forest, the use of machine vision to collect the tapping surface information of the rubber tree will be affected by unfavorable factors such as light and noise, which will lead to significant errors in the final identification and positioning. Therefore, it is necessary to enhance further the data set before network training to prevent over-fitting of the model and improve the generalization ability of the network model to adapt to the complex rubber forest environment. This study performed various random enhancement operations on the annotated original data set, including adding noise, changing light, changing pixels, translation, stitching multiple pictures, and flipping, as shown in Figure 2 . In order to ensure the balance of the proportions of various categories in the data set, a method of different enhancement times for other categories is adopted. Categories with a smaller proportion have an increase in times of enhancement, while categories with a larger proportion have a reduced number of improvements. Finally, it is divided into a training set, a verification set, and a test set in a ratio of 8:1:1. The number of pictures is 5712, 715, and 715, respectively. Table 1 shows the change in the number of category labels before and after the enhancement.

YOLO (You Only Look Once) is the beginning of the One-Stage detection algorithm. Compared with Two-Stage algorithms, YOLO can greatly improve the detection speed while ensuring good detection accuracy. According to the scale of the network, the YOLOv8 model can be divided into five versions, namely YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x, and each version includes three versions of object detection, segmentation, and key point detection. Considering the actual rubber-tapping situation, this article selected the lightweight YOLOv8n key point detection algorithm for research. Compared with the other four, YOLOv8n has a lightweight parameter structure, which is more conducive to deployment in small mobile devices.

The key point detection network structure of YOLOv8 is composed of a backbone network, a neck network, and a head network, as shown in Figure 3 . First, the input image enters the backbone network within which the CBS module, C2F module, and SPPF module are used to extract features at various scales. Then, the neck network uses the Path Aggregation Feature Pyramid Network (PAFPN) structure to process further and fuse the extracted multi-scale features. Finally, the head network processes the fused feature maps at different levels to output the detection results.

Traditional convolution uses the same parameters in each receptive field to extract feature information through the convolution kernel without considering the different information between different positions. This results in a large amount of redundant information in the extracted data, which reduces the extraction time. The efficiency of features greatly limits the performance of the model. The emergence of the spatial attention mechanism enables the model to focus on certain key features (Park et al., 2020; Li et al., 2019; Luo et al., 2022), enhancing the network’s ability to capture detailed feature information. However, it can only be used to solve the identification of spatial features and does not completely solve the parameter-sharing problem of larger convolution kernels (such as 3×3 convolution). In addition, they cannot judge the importance of each feature in the receptive field, such as the existing Convolutional Block Attention Module (CBAM) (Woo et al., 2018) and Coordinate Attention(CA) (Hou et al., 2021).

The proposal of RFA solves the limitations of the existing spatial attention mechanism and provides an innovative solution for spatial processing. Among them, the Receptive-Field Attention Convolution (RFAConv) (Zhang et al., 2023a) designed based on RFA not only emphasizes the importance of different features within the receptive field slider but also gives priority to the receptive field space features, completely solving the problem of convolution kernel parameter sharing, as shown in Figure 4 .

In RFA, the entire operation process can be divided into two parts. The first part uses group convolution to extract receptive field spatial features quickly. The second part learns the attention map by interacting with the receptive field feature information to enhance the network’s ability to extract features. However, allowing each receptive field feature to interact will incur a large computational cost. To reduce the computational cost and parameter amount as much as possible, AvgPool is first used to fuse the global information of each receptive field feature, followed by a 1×1 group convolution operation to interact with the information. Finally, the Softmax function obtains the importance of each feature in the receptive field feature. After both parts are completed, the final feature information is obtained by multiplication, as shown in Equation 1.

where, A r f and F r f represent the attention map and the transformed receptive field space feature map, respectively; g i × i and g k × k are group convolutions of size i × i and k × k , respectively; N o r m and X are normalization and input features, respectively.

The feature map obtained through RFA will not overlap the receptive fields after shape adjustment. Therefore, the learned attention map not only contains all the feature information in each receptive field but does not need to be shared in each receptive field. Finally, a standard convolution with a convolution kernel of k × k and a stride of k is used to extract feature information.

Consequently, in this paper, by replacing the standard convolutional Conv in the CBS module of the backbone network with RFAConv as depicted in Figure 5 , the feature extraction capability of the backbone network is improved, while the increase in the computational cost and the number of parameters are almost negligible.

In YOLOv8, the main task of the backbone network is feature extraction, but in detection and positioning tasks, the detected objects are multi-scale, and single-scale features cannot be used to detect multi-scale objects. Therefore, YOLOv8 uses the PAFPN structure in the neck network to process the features extracted from the backbone. Initially, the features are fused from top to bottom and then enhanced from bottom to top before generating a multi-scale feature map. Nonetheless, this approach encounters a new issue. In the process of top-down fusion, the high-level feature information may be lost or degraded, while in the bottom-up process, the low-level feature information may be lost or degraded. To address this problem, this paper references the Asymptotic Feature Pyramid Network (AFPN) (Yang et al., 2023) in the neck network, as shown in Figure 6 , to replace the original PAFPN.

As seen in Figure 6 , AFPN sequentially fuses the feature information of the bottom, middle, and top layers. This process is carried out gradually, which greatly alleviates the problem of poor feature fusion effect caused by excessive feature differences between non-adjacent layers. For example, feature fusion between the low and middle layers reduces the feature difference between them. Since the middle and high layers are adjacent layers, the feature differences between the low and high layers are also reduced.

The main task of the ASFF module in Figure 6 is to assign different spatial weights to features at various levels in the multi-level feature fusion process, which enhances the importance of key levels and reduces the impact of conflicting information between different levels. In this article, the ASFF module is divided into two modes, including ASFF2 and ASFF3. Among them, ASSF2_1 and ASSF2_2 denote level 2 feature fusion with two different weights, while ASSF3_1, ASSF3_2, and ASSF3_3 denote level 3 feature fusion with three different weights. Taking level 3 feature fusion as an example, the operation process is as follows:

where, x i j n → l represents the feature vector at position ( i , j ) from level n to level l; α i j l , β i j l , and γ i j l are the three spatial weights at level l, and the constraint is α i j l + β i j l + γ i j l = 1 ; y i j l is the feature obtained after the final fusion.

This paper evaluates the comprehensive performance of the model through two parts of experiments. The first part of the experiment: Rubber tree tapping surface detection and rubber tapping key point detection accuracy experiments, using Precision (P), Recall (R), Mean Average Precision (mAP), model parameters (Params), Flops, and FPS as evaluation indicators. Among them, P and R represent the proportion of the number of correctly predicted positive samples to the total number of predicted positive samples and the proportion of the number of correctly predicted positive samples to all positive samples, respectively; mAP is the average area under the P-R curve of all categories, which is used to measure the quality of the model in each category, among which mAP50 is the mAP value when the IOU threshold is set to 0.5; FLOPs and FPS respectively show the computing power required for model training and the inference speed of the model (the number of images inferred in 1 second).

where, TP, FP, and FN represent the number of samples correctly predicted by the model as positive (i.e., the target exists and is predicted to exist), the number of samples incorrectly predicted by the model as positive (i.e., the target does not exist but is predicted to exist), and the number of samples incorrectly predicted by the model as negative (i.e., the target exists but is predicted to not exist); P i , R i and N show the precision, recall and number of sample categories, respectively.

The second part of the experiment: Experiment on the positioning accuracy of the starting point and end point of rubber tapping, using x-axis offset distance (xOD), y-axis offset distance (yOD), and xy-axis offset distance (xyOD) as the evaluation indexes. xOD, yOD, and xyOD represent the pixel offset distances between the predicted point and the truth point in the x-axis direction, the y-axis direction, and the Euclidean direction, respectively. The calculation formulas are as follows:

where, x P , y P and x T , y T are the x-axis coordinates and y-axis coordinates of the predicted point, and the x-axis coordinates and y-axis coordinates of the truth point, respectively.

The natural rubber tree tapping surface detection and tapping key point positioning model has been improved in three parts compared to the original YOLOv8n-pose model. Part I A: The convolution RFAConv with receptive field attention mechanism replaces the ordinary convolution in the backbone network CBS module; Part II B: Neck network uses AFPN structure; Part III C: An improved key point detection head is adopted in the Head network. To verify the contribution of each improvement to the entire model, this study conducts an Ablation experiment on the natural rubber tree tapping surface detection and tapping key point positioning model. The results are shown in Table 2 .

The integration of the receptive field attention mechanism has comprehensively improved the P, R, and mAP50 of the object detection and key point detection of the original model. As shown in the experimental results of YOLOv8n-Pose+A, D_P, D_R, and D_mAP50 are improved by 1.4%, 1.2%, and 0.9%, respectively, while P_P, P_R, and P_mAP50 are improved by 1.6%, 2.1%, and 1.5%, respectively, indicating that the feature extraction capability of the backbone network has been enhanced. Meanwhile, the number of parameters and computing power cost has only increased by 0.02M and 0.3G, respectively, further proving that the receptive field attention mechanism has little impact on the size and computing cost of the entire model. After replacing the original PAFPN structure of the Neck network with AFPN, the problem of loss and degradation of bottom-level feature information and top-level feature information has been alleviated. Compared with the original model, the D_P and P_P of the YOLOv8n-Pose+B model are substantially improved by 1.9% and 2.9%, respectively. However, due to the operation of progressive fusion, feature fusion becomes more frequent, which in turn generates more parameters and computing power, increasing by 0.13M and 1.3G, respectively. After designing the original single-branch key point detection head of the Head network into a dual-branch key point detection head and introducing the residual structure, the detection head’s ability to select important features of key points is effectively enhanced. At the same time, the residual structure further compensates for the loss of important features. The P_mAP50 is improved by 2.1% compared with the original model. The combination of RFAConv, AFPN, and enhanced key point detection head showed the best detection performance, with D_mAP50 and P_mAP50 increased by 1.4% and 2.3%, respectively, compared with the original model. Still, it also increased the model complexity, increasing model size and computing power and a slower model inference speed. The biggest impact is the inference speed of the model. Although the FPS dropped to 91, the actual tapping time is 45s, and the tapping speed is about 0.8cm/s. Therefore, the detection speed of 91FPS fully meets the requirements of real-time tapping. The number of model parameters and computing power has only increased slightly, with Params increased to 3.31M and GFlops increased to 10.1G. In the current application of lightweight models in agricultural fields, Sun et al. (2022) proposed a lightweight model with 6.84M Params and 14.7G GFlops for rubber tapping, and Zhang et al. (2023b) proposed a lightweight model with 4.78M Params and 12.3G GFlops for animal recognition. In comparison, YOLOv8n-IRP is much smaller than them in parameters and GFlops, which is very beneficial for deployment on intelligent rubber-tapping machines.

Figure 9 shows the training results of the model more intuitively. Before 30 epochs, the loss of the object and key points decreases rapidly, while the mAP50 increases rapidly, indicating that the model has a faster convergence rate both before and after the improvement, and does not decrease due to the increase in model complexity. Simultaneously, the improved YOLOv8n-IRP has lower loss and higher mAP50 than the original model. This observation shows that the combination of RFAConv, AFPN, and the improved key point detection head enables the model to have better detection performance.

To demonstrate the comprehensive performance of the improved YOLOv8n-IRP model in rubber tree tapping surface detection and rubber tapping key point detection, this experiment uses three popular object detection and key point detection algorithms for comparison, as shown in Table 3 . In Table 3 , except for Faster_RCNN-RTMPose, the other three algorithms are lightweight models. Among them, the improved lightweight model YOLOv8-IRP has the highest D_P and D_R, P_P, P_R, D_mAP50 and P_mAP50, which are second only to the Faster_RCNN-RTMPose, reaching 98.5%, 88.9%, 99.2%, 89.8%, 98.3% and 86.4%, respectively. The reason why the Faster_RCNN-RTMPose can show high detection accuracy in key point detection is due to the detection mode of RTMPose. The RTMPose is a top-down key point detection algorithm. It first detects the object box and then predicts the key points in the object box by generating a key point heat map. This makes detection accuracy better than lightweight models that simultaneously predict objects and key points through regression. However, it also exposes its shortcomings. It needs to train two models: the object detection model Faster_RCNN and the key point detection model RTMPose, which makes its model size larger and requires higher computing power to train the model. The increase in model size and the cumbersome detection steps also greatly reduce the detection speed. As shown in Table 3 , the Faster_RCNN-RTMPose has the largest Params and GFlops, reaching 54.42M and 199.5G, respectively, and the smallest FPS, only 13f/s, which is extremely disadvantageous for deployment on mobile devices. On the other hand, the Params and GFlops of the YOLOv8n-IRP have only 3.31M and 10.1G, which are dozens of times smaller than the Faster_RCNN-RTMPose. At the same time, the FPS can reach 91f/s, which is several times faster than the Faster_RCNN-RTMPose, so it is more suitable for deployment on mobile devices.

The rubber forest mainly includes rubber trees with one, three, and five year(s) of tapping age. Therefore, this experiment visualizes the detection results of four models for rubber trees with one, three, and five years of harvesting age, as shown in Figure 10 . Among them, YOLOv8-IRP achieves more than 96% confidence in the detection of tapping surfaces at one, three, and five year(s), which is 2-3% higher than YOLOv5n-Pose and YOLOv8n-Pose, and it can accurately detect the presence of the starting and end point. Although compared with the Faster_RCNN-RTMPose, it fails to predict the occluded key points (the occluded key points predicted by the Faster_RCNN-RTMPose are shown in the green dotted circles in Figure 10 ), in actual rubber tapping, rubber tapping can only be carried out if the tapping key points are revealed. The occluded starting and ending points have no practical significance for rubber tapping. Therefore, the detection accuracy of the YOLOv8n-IRP meets the requirements of rubber tapping. In addition, The YOLOv5n-Pose and YOLOv8n-Pose have false detection in key point detection, which is mainly manifested in detecting key points from the tapping surface without key points, as shown in the yellow dotted circle in Figure 10 . This is because the tapping surfaces that expose key points have similar features to those that do not, while the Neck network structure of both YOLOv5n-Pose and YOLOv8n-Pose is PAFPN. The loss or degradation of low-level and high-level features will occur during the feature fusion process. Therefore, it is not possible to distinguish such similar features well enough to make correct predictions. To this end, this paper first uses the RFAConv in the YOLOv8n-IRP to enhance the ability of feature extraction. Then, it uses the AFPN structure to reduce the loss and degradation of low-level and high-level features in the feature fusion process. Finally, the designed dual-branch key point detection head is used to improve the feature screening ability and solve the problem of low prediction accuracy of similar features.

In addition to this, the uncertainty of weather and the shading of light by rubber tree trunk foliage result in variable lighting, which is one of the main challenges for vision applications in rubber forests. Therefore, in order to further demonstrate the usefulness of the improved YOLOv8n-IRP model in rubber forests, this experiment is conducted to test the overexposed, underexposed and normally exposed pictures using four models, respectively, and the detection success rates of the four models in the face of different lighting conditions are counted, as shown in Table 4 . In Table 4 , the overexposed, underexposed and normal exposure images used for testing are 200 images, respectively, in which the YOLOv8n-IRP model achieves a detection success rate of 91% in the normal exposure environment, which is more than 5% higher compared to both YOLOv5n-Pose and YOLOv8n-Pose, achieving a higher detection accuracy. For overexposure and underexposure, the detection success rates of the four models have decreased to different degrees, which is caused by 1) the insufficient number of images of complex scenes in the training set and 2) the increased difficulty of extracting important features in complex scenes, which makes the models suffer from the phenomena of misdetection and underdetection. Although the accuracy of YOLOv8n-IRP is reduced by the influence of complex illumination conditions, it still maintains an average detection success rate of 87%, which significantly improves the detection accuracy compared with the original model YOLOv8n-Pose. As shown in Figure 11 , the duplicate detection and misdetection that originally appeared in overexposure and underexposure are improved, which indicates that YOLOv8n-IRP has a more excellent feature extraction capability and enhanced robustness. While Faster_RCNN-RTMPose has a slightly higher detection accuracy than YOLOv8n-IRP in various exposure scenarios, YOLOv8n-IRP is more suitable to be deployed in mobile devices for intelligent rubber tapping use, considering the detection accuracy, model size, detection speed and the actual situation of rubber tapping.

To demonstrate the positioning accuracy of the improved YOLOv8n-IRP model at the starting and ending points of rubber tapping, this experiment calculates the xOD, yOD, and xyOD of 550 key points predicted by the four models, and their average values ​​are shown in Table 5 and Figure 12 . As can be seen from Figure 12A , the average error of the YOLOv8n-IRP on the x-axis and y-axis is lower than that of the YOLOv8n-Pose and YOLOv5n-Pose, and the accuracy has been significantly improved. It can be seen from Table 5 that the average offset error of the YOLOv8n-IRP in the x-axis direction is only 23.05 pixels, which is the smallest error among the four models; the average offset error in the y-axis and Euclidean directions is similar to that of the Faster_RCNN-RTMPose and is reduced by more than 10 pixels compared to YOLOv8n-Pose and YOLOv5n-Pose. From Figures 12B, C , it can be seen that the stability of the localization error of YOLOv8n-IRP is greatly improved compared with that of YOLOv8n-Pose, in which the maximum error does not exceed 100 pixels, while YOLOv8n-Pose shows an error of close to 180 pixels, further proving that the positioning accuracy is improved after the model improvement. Positioning accuracy and stability affects the regularity of the tapping surface, which in turn affects the efficiency of glue flow. Therefore, the improvement of YOLOv8n-IRP positioning performance has improved the efficiency of glue flow, thereby increasing the latex yield.

To further prove the feasibility of key point positioning of the YOLOv8n-IRP model, this experiment visualizes four models’ key point positioning results for rubber trees of different tapping ages, as shown in Figures 13 – 15 . Among them, the key points predicted by the YOLOv8n-IRP on rubber trees with one, three, and five year(s) of tapping age are close to the truth key points and show high positioning stability, as shown in the red dotted box in Figures 13 – 15 . However, the positioning deviation of YOLOv8n-Pose and YOLOv5n-Pose are obvious, with large error fluctuations. The Faster_RCNN-RTMPose has the lowest average offset error in the y-axis and Euclidean direction among the four models. Still, it is only a few pixels lower than the improved YOLOv8n-IRP, which is a small improvement for a 4000×6000 pixel photo. Nevertheless, in the visualization experiment, although Faster_RCNN-RTMPose achieves the highest positioning accuracy, as shown by the green dashed box in Figures 13 – 15 , there were also tapping surfaces with poor positioning, as shown by the yellow dashed box in Figures 13 – 15 , indicating that the positioning error of Faster_RCNN-RTMPose fluctuates greatly. To sum up, the YOLOv8n-IRP shows better performance in locating key points on the tapping surface, which better meets the rubber tapping requirements.