The vertical (top view) and front views of leafy orchids effectively capture the randomness and regularity of leaf morphology, respectively (Guan et al., 2011; Rodrigues et al., 2013). To obtain these complementary perspectives, a multi-view automatic acquisition device was designed for capturing images of Cymbidium goeringii (Rchb. f.) ( Figure 1A ). While the front view provides more detailed phenotypic information and thus requires a higher resolution, the proposed algorithm is capable of processing images with varying resolutions.

Two industrial cameras were employed: a Daheng industrial camera (MER2-1220-32U3C, resolution: 4024×3036) for capturing front view images, and a Hikvision industrial camera (MV-CU060-10GC, resolution: 3072×2048) for capturing top view images. To verify the generalization ability of the proposed algorithm, a publicly available dataset of individual maize plants was used, which was captured using a Grasshopper 3 camera. The parameters of the three types of cameras are shown in Table 1 . Notably, the distances between the cameras and the orchids were not fixed during data collection. Instead, these distances were dynamically adjusted based on the height and crown width of each orchid, ensuring that the entire plant was fully captured in both the top-view and front-view images.

The potted orchids were positioned at the center of a motorized turntable, which was controlled via serial communication with a multi-view image acquisition software ( Figure 1B ). The motorized turntable is controlled via RS-485 using the Modbus-RTU protocol through a USB-to-RS485 converter. Commands for absolute angle setting, step execution, and start–stop were issued from a Python 3.10 client using the pySerial library, with standard Modbus frames and a 9,600-baud 8-N-1 configuration. Module 1 captures the top view, Module 2 captures the front view, Module 3 controls the rotation angle of the motorized turntable for view selection, and Module 4 performs automated batch acquisition. This automated system enabled the turntable to adjust the viewing angle, capture images, and store data without manual intervention, thereby ensuring consistency and standardization throughout the data collection process. In total, 367 orchids with both vertical and front view images were collected during the Third China Spring Orchid Festival (Shaoxing, Zhejiang, February 22–25, 2024).

The hue channel in HSV (Hue, Saturation and Lightness) color space directly determines the color type (Hu et al., 2023; Shi et al., 2020), which facilitates accurate identification of regions of orchid leaves. The lower and upper thresholds for the green hues were established as the initial mask ( M a s k A ), where the leaf regions were highlighted in white while the rest of the background was set to black ( Figure 2A ). However, due to the presence of background elements with colors similar to the leaves and the impact of lighting variation on leaves, the binarized image displayed noticeable gaps within the leaf regions and speckled noise in the background.

To eliminate the noise and accurately extract the contours of the leaves, a series of morphological operations were applied to the M a s k A . Firstly, a morphological closing operation was used, followed by an opening operation, resulting in an improved mask M a s k A ′ . The opening and closing operation were expressed as

where ⊖ and ⊕ represents erosion and dilation operation respectively. A is the binary image, B is a 5×5 ones matrix kernel that is used to probe and interact with A . The closing operation helped bridge small gaps and holes within the leaves (Le et al., 2020), while the opening operation effectively removed speckled noice within the background (Lei et al., 2019). Next, the binarized image is further refined based on the area and shape of the leaf contours. The contours of all connected regions within the white mask were traced pixel by pixel. Due to the presence of nested contours caused by noise, only the external contours were retained. To eliminate small contours unlikely to represent leaves, contours with an area smaller than 500 pixels were filtered out. The minimum bounding rectangle was then extracted for each contour, with its aspect ratio (the length ratio of longer side to shorter side) effectively distinguishing the elongated leaves from other objects. As the length of leaves are significantly longer than the width, contours with an aspect ratio greater than 2 were preserved. These contours, forming the set S c o n t , formulating a separate mask M a s k B . The intersection of M a s k A ′ and M a s k B produced the final leaf mask M a s k l e a f , which highlighted the leaves in white against a black background, as shown in Figure 2B .

The keypoints on the orchid leaves were derived from randomly sampled points, with even distribution across the extracted leaf areas. Taking the vertical view image as an example, to simplify the computation, a circle C p o t was drawn with the centroid of M a s k l e a f as the center and half the length of the shortest contour in the set S c o n t as the radius. This circle, C p o t , generally covered the central part of the flowerpot in vertical view and served as a white mask representing the pot. As illustrated in Figure 2C , to ensure an even distribution of random points across the white regions of the leaf image, hierarchical sampling was employed to divide the segmented leaf area into 40×40 patches. A random point is sampled from each subregion, excluding those within C p o t . Subsequently, morphological erosion algorithm is applied to reduce the boundary regions of the leaves (Yin et al., 2023), preventing points near the edges of the white areas from being selected as keypoints and ensuring that the generated random points were located close to the central skeleton of the leaves.

For both random and regular morphological leaves, the outermost keypoints were firstly identified, then the search direction for subsequent keypoints was determined to reduce spatial complexity. The traversal of keypoints for the next leaf begins only after all keypoints of the current leaf have been identified. For each leaf, P i j represents the i -th keypoints of the j -th leaf. P 1 1 was designated as the first keypoints of the first leaf (marked as visited), which has the maximum Euclidean distance from point C p o t (closest to leaf tip), as illustrated in Figure 3A . P 2 1 was identified as the closest unvisited point to P 1 1 . To locate P 2 1 , a circular search area is iteratively expanded around P 1 1 until one or more unvisited points were found. If a single unvisited point is identified within the search area, it is designated as P 2 1 . If multiple unvisited points were found, the point closest to P 1 1 is selected as P 2 1 , which is then marked as visited, forming the vector P 1 1 P 2 1 → .

The search for P 3 1 is also based on extending the circular search area centered on P 2 1 . However, given that the orientation of the leaf skeleton is determined, half of the random points within the circular search area are not candidate points for P 3 1 . The direction of the skeleton informs the search direction for subsequent keypoints, thereby reducing the search space for the next keypoint. The skeletal direction is determined by the relative positions of points P 1 1 and P 2 1 , specifically by comparing the absolute differences between the horizontal coordinates ( P 1 x 1 and P 2 x 1 ) and the vertical coordinates ( P 1 y 1 and P 2 y 1 ). This will result in more efficient search. The rules for determining the search direction for keypoints were shown in Table 2 .

In vertical view of orchids, the uncertain growth trajectory of each leaf results in a random morphology. Therefore, based on the predetermined keypoint search direction, the optimal keypoints closest to the central skeleton must be adaptively identified. The determination of P 3 1 is based on the angle formed between vectors P 1 1 P 2 1 → and P 2 1 P 3 1 → , formulated as

where P 3 ( i ) 1 is the i -th candidate point of P 3 1 in circular search area. θ P 1 1 P 2 1 P 3 ( i ) 1 is the angle between P 1 1 P 2 1 and P 2 1 P 3 1 . Within the left semicircle of the circular search area, there are three candidate keypoints, resulting in three angles. In order to effectively capture the natural curvature characteristics of orchid leaves along their main skeleton, we set a fixed threshold for the angular difference of candidate keypoints in our algorithm. Through statistical analysis and experimental validation on multiple orchid samples, we found that the local curvature variations of most leaves are confined within a narrow range. When the angular difference between a candidate keypoint and the current skeleton direction is less than 20°, the true turning points can be effectively identified while avoiding interference from noise and local anomalies. If the minimum angle in S 1 θ is less than 20°, the corresponding candidate point is selected as P 3 1 . Otherwise, the circular search area is further expanded. This 20° threshold was chosen based on the observed morphological properties of orchid leaves and extensive empirical testing, which demonstrated that it provides a robust balance between sensitivity (capturing genuine turning points) and specificity (avoiding spurious points due to noise). Fixing this threshold not only reflects the inherent geometric properties of orchid leaves but also simplifies the algorithm structure, thereby enhancing computational efficiency and consistency.

Once the first three keypoints have been identified, the subsequent keypoints were determined by iteratively running the same search algorithm. For example, to determine P 4 1 , the circular search area is iteratively expanded with P 3 1 as the center. The unvisited points within the left semicircle of this area form a set P 4 ( 1 ) 1 , P 4 ( 2 ) 1 , … , P 4 ( N ) 1 . The algorithm then calculates the angle set S θ = { θ ^ i − 1 ( k ) | θ ^ i − 1 ( k ) = ∠ P 3 1 P 2 1 P 4 ( k ) 1 , k = 1 , 2 , … , N } and θ ^ i − 2 = ∠ P 2 1 P 1 1 P 3 1 . It compares the differences between θ ^ i − 2 and each θ ^ i − 1 ( k ) . Similar to the selection of P 3 1 , if the smallest angle difference exceeds the threshold of 20°, the circular search area is further expanded. Otherwise, the point with the smallest angle difference was selected as P 4 1 , as follows

Then, the same method was iteratively applied to locate subsequent keypoints P i 1 . The loop terminates under the following condition

where P l a s t represents the last identified keypoints of the leaf, r p o t is the radius of C p o t , r l a s t is the radius of the circular search area corresponding to P l a s t , and | | C p o t − P l a s t | | is the Euclidean distance between the centers of C p o t and P l a s t . When the circular search area of P i 1 contains no unvisited points and intersects with C p o t , P i 1 is designated as P l a s t , and the iteration stops. At this time, all keypoints for the single leaf have been determined.

The algorithm is then repeated on other leaves. The point P 1 2 , which has the greatest Euclidean distance from C p o t , is identified as the first keypoints of the second leaf. The same algorithm was then applied to determine all keypoints of this leaf, continuing until the keypoints for all leaves were found.

In front view of orchids, due to the influence of gravity, all orchid leaves form a completely regular convex polyline ( Figure 3B ). Therefore, the trajectory pattern was fitted by minimizing curvature based on the consistent leaves trend. Let any three continuous keypoints along the leaf (from the root to the tip) be denoted as P i − 1 1 , P i 1 , and P i + 1 1 . These points must satisfy the following condition

where N j represents the number of keypoints on the j -th leaf. This ensures that any adjacent three keypoints form a convex sub-polyline, and the collection of these sub-polylines constitutes the fully convex skeleton of the leaf.

The keypoints identification process for the front view image of the orchid leaf begins by selecting the initial point P 1 1 as the farthest point from the leaf base. To determine the subsequent point P 2 1 , a circular search area centered on P 1 1 is iteratively expanded until more than two unvisited points were found. The point with the closest y -coordinate to P 1 y 1 is then selected as P 2 1 . The next keypoint, P 3 1 , is identified by expanding the search area centered on P 2 1 . If more than 2 unvisited points were found within this area, the point that forms a convex curve with P 1 1 and P 2 1 and results in minimal curvature is chosen as P tem 1 1 .

The search then continues by expanding the area around P tem 1 1 . If there were at least two points in this area, they were evaluated. If a point P tem 2 1 forms a convex polyline with P 1 1 and P 2 1 and lies above P tem 1 1 , this indicates that the polyline P 1 1 P 2 1 P tem 2 1 has a lower curvature than P 1 1 P 2 1 P tem 1 1 . In this case, the following vector addition is performed as

P 2 1 P t e m 1 1 → + P t e m 1 1 P t e m 2 1 → = P 2 1 P t e m 2 1 →

Thus, P tem 2 1 is set as the final P 3 1 , and P tem 1 1 is discarded. Conversely, if P tem 2 1 is located below P tem 1 1 , then P tem 1 1 is confirmed as P 3 1 , and P tem 2 1 becomes P 4 1 . This iterative traversal, combined with vector addition, allows for a more precise fitting of the leaf structure in front view.

All experiments—orchid (two industrial cameras) and maize (PHENOARCH platform)—were run with a single, immutable parameter set. Values were selected once on an orchid batch and were not tuned thereafter, thereby demonstrating cross-species generalizability. Table 3 lists the constants, and no entry changes across datasets, resolutions or cameras.

The accuracy of the predicted skeletons was evaluated using curvature error ( Figure 4 ). Let y ( x ) represent the true skeleton curve of the leaf, with curvature at point x i given by (Schamberger et al., 2023)

and the curvature at point x i of the predicted leaf skeleton polyline is given by (Chuon et al., 2011)

where θ i represents the angle at the i -th vertex of the polyline, and d i − 1 , i is the Euclidean distance between the ( i − 1 ) -th and i -th points on the polyline. The curvature error between the predicted and true skeletons is defined as

A smaller curvature error indicates a higher shape conformity between predicted polyline skeleton and the true curve skeleton. The leaf recall rate was used to measure the proportion of leaves correctly identified and successfully skeletonized out of the total number of leaves, defined as

For vertical view, the number of algorithm’s outermost loop (which traverses each leaf) corresponds to the leaves’ number. As shown in Figure 5A , since the initial point P 1 j for each leaf is closest to the leaf tip, the total number of P 1 j represents the leaf count. The maximum distance between any two P 1 j represents the crown width of the orchid (Ebrahimi et al., 2020), measured in pixels. Because pixel measurements cannot accurately reflect the true crown width, the relative crown width was used as a key feature for evaluating the orchid, formulated as

Let N j represent the number of keypoints in the j t h leaf. Since the skeleton formed by connecting these keypoints accurately represents the shape of the leaf, the total length of the skeleton corresponds to the length of the leaf. Leaf height consistency is the variance of each leaf’s mean vertical distance from keypoints to the pot’s top reference line in the front view, reflecting canopy regularity that growers use to judge form. Relative crown width is scale robust by construction. Leaf height consistency is expressed as

A concentric circle C c r o w n is drawn with the crown width as its diameter, centered on the flowerpot ( Figures 5A, B ). Between C c r o w n and C p o t , N additional concentric circles C j ( j = 1 , 2 , … , N ) were generated with equal radial increments, where the radius of each C j is denoted as R C j . The boundaries of each concentric circle may intersect the skeleton of the leaves at different points. Each intersection divides C j into several arcs, with l i j representing the arc length of the i -th segment of the j -th concentric circle, corresponding to the distance between adjacent leaves under the condition of R C j . Thus, V a r ( l j ) represents the leaf distribution consistency at R C j .

However, in vertical view, the leaves radiate outward from the center, causing the distance between adjacent leaves to vary under different R C j conditions. Generally, l i j is positively correlated with R C j , and leaf distribution consistency at an intermediate R C j better reflects the overall leaf uniformity of the orchid. Therefore, leaf distribution consistency for the orchid (the weighted average of V a r ( l j ) ) quantified how uniformly leaves occupy space around the pot by partitioning the annulus between the pot circle and the crown circle into concentric rings. Leaf distribution consistency is defined as

where λ j is a Gaussian function satisfying ∑ j = 1 N λ j = 1 and a r g m a x { λ j } = N 2 , expressed as

For front view, the top edge of the flowerpot’s rectangular boundary serves as the bottom reference line of leaves ( l r ). The distance from the i -th key point of leaf j to l r is denoted as h i j , and the average distance of all keypoints in leaf j to l r is represented by h ^ j = 1 M j ∑ k = 1 M j h k . The variance of h ^ j across all leaves reflects the variation in the average height of each leaf within the orchid, which is defined as leaf height consistency ( Figure 5C ), indicating whether the architecture is harmonized rather than a mix of very long and very short leaves. Leaf height consistency is formulated as

where M j represents the number of keypoints on leaf j in front view.

The keypoints connection method was applied to both front view and vertical view images of the orchid. The keypoints connection result was shown in Figures 6A, B , respectively It can be seen that the leaves were accurately extracted without any background noise. Due to the application of morphological erosion to reduce the leaf boundary regions, random points were uniformly generated in areas close to the central skeleton. Since the leaf tip width is narrow, the diameter of the random points exceeded the eroded width of leaf tip, resulting in incomplete point generation at the outermost tip. After executing the keypoints algorithm, nearly all leaf skeletons were accurately extracted under complex conditions involving dense leaves, leaf intersection and leaf occlusion. Motivated by this observation, a quantitative analysis was conducted to explicitly characterize when tip detection fails and how this affects downstream phenotypes.

To make failure modes explicit, a leaf-level criterion was adopted whereby a tip is counted as missed if the outermost 20% sector of that leaf contains no sampled points. Using this definition and expressing the random-point marker diameter as a percentage of the shortest contour length in S c o n t , the tip-miss frequency over the entire orchid dataset (front and top views combined) was 19.24% at the default setting of 1.5%. An ablation varying the marker diameter showed a monotonic decrease in misses—from 32.34% (6%) to 0% (0.1%)—indicating that erosion-induced narrowing at very thin tips is the primary cause and can be mitigated by smaller markers. Phenotype level effects were limited at the default: leaf count was essentially unchanged. Relative crown width exhibited a mild negative bias (−2.4%). Leaf-length consistency showed a small positive bias (+2.6%). Leaf distribution consistency changed negligibly due to Gaussian mid-radius weighting. And in front views, leaf-height consistency showed a minor negative bias (−1.1%). These results indicate that, under default settings, tip misses occur in a fraction of leaves but have minor impact on the five target phenotypes. If complete tip recovery is required, reducing the marker diameter (≤0.3%) eliminates the effect without altering the rest of the pipeline.

YOLOv7-SlimPose (Gao et al., 2024) was utilized based on the same dataset. For each leaf, four keypoints were predefined and annotated ( Figures 4A, B ): (1) Root point, which is the point where the base of the plant connects to the pot; (2) Angle point, located at the midpoint between the Root point and the Tip point; (3) Angle point 2, positioned at three-quarters of the distance from the Root point to the Tip point; and (4) Tip point, representing the apex of the leaf. Keypoints point (1) through (4) were sequentially connected to form the leaf skeleton. 10 leaves and 8 leaves were annotated in vertical view images and font view images respectively. The annotated leaf skeletons in each image were evenly distributed. A total of 561 images were selected for the training set, 126 images for the validation set, and the remaining 15 images were used as the test set. The ratio of vertical view to front view images was maintained at 1:1 across the training, validation, and test sets. Due to the distinct distribution of leaves in the vertical and front view, the images from these two perspectives were separated and used to train two models: one for vertical view and another for front view (Total training time is 7.3 hours). After training, the test set images were input into YOLOv7-SlimPose, which output the leaf keypoints and skeletons for front view (blue output in Figure 4C ) and vertical view (blue output in Figure 4D ).

The test set images ( Figure 7A ) were processed by both algorithms. Based on the identified keypoints and the leaf skeleton, 5 key features were further extracted, and the curvature error and leaf recall rate were calculated for each, as shown in ( Figure 7B ).

YOLOv7-SlimPose has a fixed output layer structure and relies on an anchor-based object detection mechanism, which restricts the size and position of the output targets within a predefined range (Liu et al., 2022). The model generates a fixed number of keypoints for each anchor position during forward propagation. This design limits the model to outputting only a predetermined number of keypoints and predefined skeleton connections (He et al., 2022; Tan et al., 2024). Consequently, skeletons’ number and keypoints detected by YOLOv7-SlimPose model are entirely dependent on the annotations in the training data, which often require expert knowledge in botany, especially when dealing with complex plant structures.

For front view, orchid leaf root points were distributed across various positions within the pot. However, due to annotation constraints, YOLOv7-SlimPose identifies all root points as being concentrated in a single location, failing to represent the true morphology of leaves. Orchids exhibit varying numbers of leaves, leaf lengths, and leaf distributions, but YOLOv7-SlimPose lacks the flexibility to dynamically adjust the output structure based on the input, making it inadequate for handling uncertain keypoints numbers and connection patterns. Therefore, deep learning models like YOLOv7-SlimPose are more suitable for applications involving fixed keypoints numbers and skeleton structures, such as human or animal pose estimation, rather than for complex plants like orchids, which have dense and variable leaf pattern.

Based on the skeletons generated by YOLOv7-SlimPose, five orchid phenotypes were extracted from the test set, with the results shown in Figure 7 . For the vertical view, as the annotated number of leaves was fixed at 12, the model consistently detected 12 leaf skeletons (with the leaf count remaining constant at 12) and 37 keypoints across all images. In contrast, the method proposed in this work accurately detected varying numbers of leaves, achieving a leaf recall rate of 1 in three out of five samples. In the first two samples, where the leaves were relatively sparse, YOLOv7-SlimPose was able to detect most of the leaf tips, resulting in a relative crown width close to the ground truth. However, as the number of leaves increased in the remaining samples, the accuracy of the crown width predictions by YOLOv7-SlimPose dropped significantly compared to this work. The fixed number of keypoints output by YOLOv7-SlimPose resulted in poor fitting of the true leaf skeleton curve, limiting the model’s ability to accurately reflect leaf consistency. The average curvature error and leaf recall on test set were only 59% and 63%, respectively. In contrast, this work successfully detected enough leaves and keypoints, and accurately fitted the real leaf skeleton, resulting in more precise extraction of leaf length consistency, distribution consistency, and height consistency. Without manual labeling and long-term training, the average curvature error and blade recall of the test set reached 0.12 and 92%, respectively.

To demonstrate how the proposed spontaneous key-points connection differs from traditional non-learning approaches, the same binary masks were processed with the medial-axis transform (MAT) and with Zhang–Suen thinning ( Figure 8 ). Both classical algorithms inherit every imperfection present in the thresholded image: darker leaf segments translate into complete breaks, while small contour irregularities give rise to dense clusters of spurs and lateral branches. The resulting skeletons are fragmented and highly branched, preventing the continuous center-line that downstream phenotype measurements require. By contrast, the proposed strategy links interior sample points under geometric constraints that actively suppress burr formation and bridge minor gaps, yielding a single, smooth curve for each leaf and enabling reliable extraction of length, curvature and spatial-distribution traits. The qualitative comparison therefore reinforces the practical advantage and uniqueness of the present method for complex leafy plants.

Experiments were conducted on a CPU-only workstation equipped with 32 GB DDR4–2666 system memory. For images of 4024 × 3036 and 3072 × 2048 pixels, the algorithm required on average 3.556 s and 2.247 s per image, respectively, with peak resident memory of approximately 190.50 MB and 150.37 MB. No GPU acceleration was used.

Compared with the classical methods that have been publicly disclosed, iterative refinement in the i5-4670K and 4 GB memory virtual machine environment, most algorithms are completed within the range of 0.58-0.96 s for a single image, while Stentiford and an improved method take about 1.21-1.43 s (Gramblicka and Vasky, 2016). The total runtime given in the example script, represented by the central axis transformation, is about 1.34 s/image. The total runtime of the activity outline in similar official examples is about 2.03 seconds per image. The Hessian Frangi ridge filter shows an average of 1.187 s/graph in the i7 3.0 GHz, 64 GB, RTX 2060 environment, and hardware configuration is provided to ensure comparability (Hachaj and Piekarczyk, 2023).

Although the single frame time is slightly higher than several classic skeleton pipelines, this method does not require labeling and training, and can adaptively generate a variable number of keypoints and skeletons after one preprocessing. It directly supports phenotype calculations such as leaf number, crown width, length consistency, distribution consistency, and height consistency, and has been validated for stable leaf recall rate and cross species generalization in orchid and maize morphologies. The practicality of this type of “training species independent” is supported by the cost of running a CPU in seconds and a hundred megabits of memory, avoiding the time and economic expenses of retraining and manual annotation. Regarding scalability, when the number of pixels increases from 6.29 million to 12.22 million (1.94 ×), the average time takes from 2.247 s to 3.556 s (1.58 ×), showing a nearly linear or even better scaling. The algorithm mainly relies on local search and angle/convexity constraints on the mask, and the computational complexity mainly increases linearly with the number of pixels and candidate points, so it will only be faster at lower resolutions. 4024 × 3036 has significantly exceeded the imaging requirements of most production processes, and high-resolution is chosen to cover the most demanding resolution scenes. In terms of density, the high leaf density, occlusion, and overlap of orchids have constituted strict stress tests, while the full growth period samples of corn gradually increase leaf density, all of which can maintain stable skeleton extraction and high leaf recall rate, indicating that the scalability in leaf quantity and density has practical significance.

To verify the generalization ability of the proposed algorithm, it was applied to maize. Similar to orchid leaves, maize leaves are soft, numerous, densely packed, and exhibit significant overlap. Using the publicly available single-plant maize dataset (Dataset URL: https://datasetninja.com/maize-whole-plant-image-dataset), which captured images of a single maize plant over 113 days—with 1 vertical view and 12 front view images taken each day—the spontaneous keypoints connection algorithm was applied to both a front view ( Figure 9 ) and a vertical view image ( Figure 9 ) from 10 different days, spanning pre-, mid-, and post-growth stages ( Figure 10 ). Due to the relatively wide leaves of maize, random points could still be generated at the eroded leaf tips, and in some cases, a single maize leaf may be mistakenly identified as two separate leaves in the vertical view.

It is important to note that the extraction of maize leaf count is based on the front view images, while the vertical view is primarily used to extract features such as crown width, leaf distribution consistency, and overall plant structure. Therefore, any misidentification of leaf number in the vertical view does not affect the final feature extraction or the evaluation of the algorithm’s generalization ability on maize leaves.

Samples #4 and #5 were annotated for YOLOv7-SlimPose training, with 6 skeletons annotated per front view image and 4 skeletons per vertical view image. The relative crown width was extracted from 10 samples ( Figure 10 ). During the early growth stage, when maize leaves are smaller and fewer in number, YOLOv7-SlimPose predicted more skeletons than the actual leaf count, allowing it to accurately estimate the crown width in this stage. However, as the plant continued to grow, both the size and number of leaves increased. The number of leaves exceeded the prediction range of YOLOv7-SlimPose in mid-to-late growth stages, resulting in greater prediction error for the relative crown width.

In contrast, the proposed method accurately predicted the relative crown width across all growth stages ( Figure 10 ) and consistently detected all leaves from the front view, maintaining a leaf recall rate of 100% throughout ( Figure 10 ). Additionally, compared to YOLOv7-SlimPose, the proposed method maintained a lower curvature error ( Figure 10 ) for front view images across all growth stages. These results confirm that the proposed method accurately extracts all keypoints and skeletons of maize leaves, demonstrating its effectiveness in keypoint detection, skeleton recognition, and feature extraction in crops with complex morphologies.

In summary, this work identifies keypoints and skeletons based solely on the connection of randomly generated points in the leaf region. This approach eliminates the need for predefined keypoints and training based on plant structure, making it adaptable for a broader range of plant phenotypic extraction tasks.