Rose maturity identification method, rose picking robot and picking method

Abstract

The invention relates to the technical field of rose recognition and picking, in particular to a rose maturity recognition method based on a YOLOv11n-FE model, a rose picking robot and a rose picking method. Comprising the following steps: acquiring rose images to form a data set; based on the improved YOLOv11n-FE model, the data set is trained, and a trained rose maturity detection model is obtained; and S2, inputting a to-be-detected rose image into the rose maturity detection model obtained after training in the step S2, and carrying out target rose maturity detection to obtain a final detection result. Accurate identification of the maturity of the roses is achieved, and cutting, picking and collecting of the identified mature roses are achieved.

Description

Technical Field

[0001] This invention relates to the field of rose identification and harvesting technology, and in particular to a rose maturity identification method based on the YOLOv11n-FE model, a rose harvesting robot, and a rose harvesting method. Background Technology

[0002] The global flower industry is booming, and the demand for roses continues to grow. However, traditional manual harvesting methods have significant drawbacks. This labor-intensive method faces challenges such as labor shortages and rising costs. Furthermore, harvesting efficiency is low, averaging only 200-400 roses per hour. Workers are easily injured by thorns, and pesticide residues pose a health threat. In addition, the accuracy of manually judging rose maturity is only about 50%-75%, affecting product quality. While existing mechanical harvesting equipment can increase speed, it is difficult to accurately identify the ripeness. Some intelligent harvesting equipment uses machine vision technology, but in complex field environments such as uneven lighting and shading by branches and leaves, it suffers from problems such as slow recognition, poor accuracy, and high cost. Deep learning has significant advantages in image recognition and object detection. When applied to rose-harvesting robots, it can accurately identify the ripeness, position, and posture of roses. However, the application of this technology is still in its early stages, facing challenges such as low recognition rates in complex environments, the need for optimization of motion control, and insufficient power consumption. Summary of the Invention

[0003] The purpose of this invention is to overcome the above-mentioned defects in the existing technology and to propose a rose maturity identification method, a rose picking robot and a picking method, which realizes the accurate identification of rose maturity and enables the cutting, picking and collection of identified mature roses.

[0004] The technical solution of this invention is: a method for identifying the maturity of roses, comprising the following steps: S1. Obtain images of roses to form a dataset; S2. Based on the improved YOLOv11n-FE model, the dataset is trained to obtain the trained rose maturity detection model; S3. Input the rose image to be detected into the rose maturity detection model obtained after training in step S2, perform target rose maturity detection, and obtain the final detection result.

[0005] In this invention, the improved YOLOv11n-FE model includes a backbone network, a neck network, and a detection head.

[0006] The backbone network consists of four detection phases, each of which includes a FasterNet-Pconv convolutional module and several stacked c3k2-FsterNet modules; The first detection stage includes a FasterNet-Pconv convolutional module and three stacked c3k2-FsterNet modules: the input image for the first detection stage is a 640×640×3 image, which is passed through the 3×3 dimensionality reduction convolution of the FasterNet-Pconv convolutional module and the three stacked c3k2-FasterNet modules, and outputs a feature map with a size of 320×320×64. The second detection stage includes a FasterNet-Pconv convolutional module and six stacked c3k2-FsterNet modules: the input feature map of size 320×320×64 is passed through the 3×3 dimensionality reduction convolution of the FasterNet-Pconv convolutional module and the six stacked c3k2-FasterNet modules, and the output feature map of size 160×160×128 is passed. The third detection stage includes a FasterNet-Pconv convolutional module and six stacked c3k2-FsterNet modules: the input feature map of size 160×160×128 is passed through the 3×3 dimensionality reduction convolution of the FasterNet-Pconv convolutional module and the six stacked c3k2-FasterNet modules, and the output feature map of size 80×80×256 is passed. The fourth detection stage includes a FasterNet-Pconv convolutional module, three stacked c3k2-FsterNet modules, an SPPF module, and a C2PSA module. The input feature map of size 80×80×256 is passed through the 3×3 dimensionality reduction convolution of the FasterNet-Pconv convolutional module and the three stacked c3k2-FasterNet modules to output a feature map of size 40×40×512.

[0007] In the neck network, firstly, the 40×40×512 feature map output from the fourth detection stage is upsampled to 80×80 and then concatenated with the 80×80×256 feature map output from the third detection stage in the first concat connection module to obtain an 80×80×768 feature map. This feature map is then passed through the c3k2 module with EMA to perform attention weighting on the fused feature map to highlight maturity features. The processed 80×80×768 feature map is upsampled to 160×160, and then spliced ​​and fused with the 160×160×128 feature map output from the second detection stage in the second concat connection module to obtain a 160×160×896 feature map. This feature map is then used by the c3k2 module with EMA to perform attention weighting on the fused feature map to highlight maturity features. The processed 160×160×896 feature map is upsampled to 320×320, and then spliced ​​and fused with the 320×320×64 feature map output from the first detection stage in the third concat connection module to obtain a 320×320×960 feature map. This feature map is then used by the c3k2 module with EMA to perform attention weighting on the fused feature map. The output 320×320×960 feature map is mainly used for small target detection, realizing the detection of immature roses. The output 320×320×960 feature map is input into the FasterNet-Pconv convolution module to adjust the channels, and then concatenated by the fourth Concat connection module before being input into the c3k2 module. The 160×160×480 feature map obtained after processing by the c3k2 module is mainly used for mid-target detection, realizing the detection of semi-ripe roses. The 160×160×480 feature map is input into the EMA module. After attention weighting by the EMA module, the feature map is input into the FasterNet-Pconv convolution module to adjust the channels. In the fifth Concat connection module, it is concatenated and fused with the 40×40×512 feature map output from the fourth detection stage. The fused feature map is then passed through the c3k2 module of the EMA module for attention weighting. The output 80×80×240 feature map is mainly used for large object detection, realizing the detection of mature roses.

[0008] This application also discloses a rose-picking robot, including a body and a collection basket at the rear of the body; The top of the vehicle body is equipped with a robotic arm telescopic mechanism. One end of the robotic arm telescopic mechanism is rotatably connected to the vehicle body, and the other end of the robotic arm telescopic mechanism is equipped with a flexible rose flower holder mechanism and a cutting and clamping mechanism. It also includes a control mechanism, with the robotic arm telescopic mechanism, the flexible rose petal holder mechanism, and the cutting and clamping mechanism all connected to the control mechanism.

[0009] The robotic arm telescopic mechanism includes: Rotational joint; The telescopic part includes a fixed branch pipe and a telescopic branch pipe. One end of the fixed branch pipe is connected to the top of the vehicle body via a rotating joint. The other end of the fixed branch pipe is slidably connected to the end of the telescopic branch pipe. The other end of the telescopic branch pipe is connected to the cutting clamping mechanism.

[0010] The cutting and clamping mechanism includes: The first connecting frame has one end fixedly connected to the free end of the robotic arm telescopic mechanism, and the other end rotatably connected to the second connecting frame. The second connecting frame is fixed with a support plate, and the support plate is fixedly connected to the cutting part and the clamping part. The clamping part includes a first clamp and a second clamp, which rotate in opposite directions, and both the first clamp and the second clamp are provided with arc-shaped elastic clamping surfaces on their opposing surfaces. The cutting section includes a first cutter and a second cutter. The first cutter is fixedly connected to a first clamp, and the second cutter is fixedly connected to a second clamp. Cutting blades are provided on the opposite surfaces of the first cutter and the second cutter.

[0011] The flexible rose petal support mechanism includes: A connecting plate is fixedly connected to a support plate, and a cutting clamping recognition camera is provided at the fixed connection between the connecting plate and the support plate; The free end of the connecting plate has two symmetrically arranged support plates with a gap between them, which support the rose holder.

[0012] This application also discloses a method for harvesting roses using the aforementioned rose-harvesting robot, comprising the following steps: A1. The rose-picking robot moves between the rows of rose bushes and collects images of the rose bushes until it collects an image that has the characteristics of the target to be picked. A2. The improved YOLOv11n-FE model identifies roses on rose plants based on the collected images and outputs the rose maturity recognition results, including the bounding box of each rose, the maturity of the rose, and the confidence score. A3. Based on the rose maturity recognition results from the improved YOLOv11n-FE model, the identified mature roses are harvested: the robotic arm telescopic mechanism moves to support the flexible rose receptacle mechanism on the receptacle of the mature rose, the cutting and clamping mechanism completes the clamping, fixing and cutting of the mature rose stem, and collects the cut mature roses. A4. Once there are no more roses to pick in the field of vision, continue searching for rose plants and repeat steps A1 to A3 until all mature roses within the work area have been picked.

[0013] The beneficial effects of this invention are: (1) The improved YOLOv11n-FE model has higher recognition accuracy and more accurate judgment of the maturity of roses: The backbone network includes four detection stages, which are used to extract low-level features (petal edges / texture), mid-level features (bud outline), high-level semantic features (bud shape) and abstract maturity features (petal color depth); The neck network strengthens maturity-related features (such as the color of mature roses and the fullness of buds) through three upsampling splicing and EMA attention weighting, and finally outputs three multi-scale feature maps to adapt to the detection needs of immature small roses (small targets), semi-mature roses (medium targets) and mature roses (large targets), avoiding missed detection or misjudgment due to differences in roses; (2) Experimental verification of the improved YOLOv11n-FE model on the rose maturity dataset shows that its mean precision (mAP@0.5) reaches 93.8%, which is 2.1% higher than the basic YOLOv11n model and has significant advantages over the existing mainstream models; the recall rate (R) reaches 0.98, which means that the model can capture mature roses more comprehensively, with fewer missed detections, and the proportion of immature roses misclassified as mature in the identification results is lower, which directly ensures the quality of picking; (3) The improved YOLOv11n-FE model has fewer parameters, higher computational efficiency, and is compatible with low-performance devices; (4) The rose picking robot proposed in this application can collect images of rose plants within the visual range and accurately judge the maturity of roses using the YOLOv11n-FE model. Then, it realizes the automatic cutting, support and collection of mature roses, realizing the entire process of automatic identification and collection of mature roses, which greatly improves the efficiency of rose picking. Attached Figure Description

[0014] Figure 1 This is a structural diagram of the improved YOLOv11n-FE model; Figure 2 This is a three-dimensional structural diagram of the rose-picking robot described in this invention; Figure 3 This is a schematic diagram of the front view structure of the rose-picking robot described in this invention; Figure 4 This is a schematic diagram of the telescopic structure of the robotic arm; Figure 5 This is a structural diagram of the flexible rose petal support mechanism and the cutting and clamping mechanism; Figure 6 This is a flowchart of the harvesting method described in this invention.

[0015] In the diagram: 1. Fill light; 2. Vehicle body; 3. Collection basket; 4. Flexible rose holder mechanism; 5. Cutting and clamping mechanism; 501. First connecting frame; 502. Second connecting frame; 503. Support frame; 504. First clamp; 505. First cutter; 506. Second clamp; 507. Second cutter; 508. Support plate; 6. Track walking mechanism; 601. Track drive gear; 7. Robotic arm telescopic mechanism; 701. Telescopic part; 8. Control system; 9. Global camera; 10. Track drive motor; 11. Camera navigation camera; 12. Power supply battery; 13. Camera navigation fill light; 14. Cutting and clamping recognition camera; 15. High torque motor. Detailed Implementation

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0017] Specific details are set forth in the following description to provide a full understanding of the invention. However, the invention can be practiced in many ways other than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0018] This application proposes a method for identifying the maturity of roses, which includes the following steps, the flowchart of which is shown below. Figure 1 As shown.

[0019] The first step is to acquire rose images and use a large number of rose images to build a training set for the model.

[0020] To ensure the authenticity and reliability of the data, the images used in this study's rose plant dataset were collected in person at a flower cultivation base in Qingdao City, Shandong Province. The original images had a resolution of 4096 pixels × 3072 pixels, were in JPG format, had an aperture of f / 1.9, and an equivalent focal length of 35 mm. A total of 2000 rose sample images were collected. To enrich the diversity of the dataset, the collected rose sample images underwent data enhancement processing including random brightness, random contrast, random cropping, and random rotation. All enhanced images were adjusted to a uniform size of 640 × 640 pixels.

[0021] The second step is to construct a rose maturity detection model based on the improved YOLOv11n-FE model.

[0022] This application utilizes the improved YOLOv11n-FE as the basic framework to construct a rose maturity detection model. For example... Figure 2 As shown, the improved YOLOv11n-FE model includes a backbone network, a neck network, and a detection head.

[0023] The backbone network is responsible for extracting multi-scale features from the input image. The backbone network consists of four detection stages, each of which includes a FasterNet-Pconv convolutional module and several stacked c3k2-FsterNet modules.

[0024] In this embodiment, the first detection stage includes a FasterNet-Pconv convolutional module and three stacked c3k2-FasterNet modules. The input image for the first detection stage is a 640×640×3 image. Through the 3×3 dimensionality reduction convolution of the FasterNet-Pconv convolutional module and the three stacked c3k2-FasterNet modules, a feature map of size 320×320×64 is output in tensor format. The output feature map stores low-level features, such as petal edges and textures.

[0025] The second detection stage includes a FasterNet-Pconv convolutional module and six stacked c3k2-FsterNet modules. The first detection stage outputs a feature map of size 320×320×64. Through the 3×3 dimensionality reduction convolution of the FasterNet-Pconv convolutional module and the six stacked c3k2-FasterNet modules, a feature map of size 160×160×128 is output in tensor format. The output feature map stores mid-level features, such as flower bud outlines.

[0026] The third detection stage includes a FasterNet-Pconv convolutional module and six stacked c3k2-FasterNet modules. The second detection stage outputs a feature map of size 160×160×128. Through the 3×3 dimensionality reduction convolution of the FasterNet-Pconv convolutional module and the six stacked c3k2-FasterNet modules, a feature map of size 80×80×256 is output in tensor format. The output feature map stores high-level semantic features, such as the overall shape of the flower bud.

[0027] The fourth detection stage includes a FasterNet-Pconv convolutional module, three stacked c3k2-FsterNet modules, an SPPF module, and a C2PSA module. The third detection stage outputs a feature map of size 80×80×256. Through the 3×3 dimensionality reduction convolution of the FasterNet-Pconv convolutional module and the three stacked c3k2-FasterNet modules, a feature map of size 40×40×512 is output in tensor format. The output feature map contains the most abstract maturity-related features, such as petal color depth.

[0028] The neck network is primarily used for feature fusion, highlighting maturity features. In this stage, firstly, the 40×40×512 feature map output from the fourth detection stage is upsampled to 80×80, and then concatenated with the 80×80×256 feature map output from the third detection stage in the first concat connection module to obtain an 80×80×768 feature map. This feature map is then passed through a c3k2 module containing EMA to apply attention weighting to the fused feature map, highlighting maturity features.

[0029] Then, the processed 80×80×768 feature map is upsampled to 160×160, and subsequently concatenated and fused with the 160×160×128 feature map output from the second detection stage in the second concat connection module to obtain a 160×160×896 feature map. This feature map is then passed through a c3k2 module containing EMA to apply attention weighting to the fused feature map, highlighting maturity features.

[0030] Next, the processed 160×160×896 feature map is upsampled to 320×320, and then concatenated with the 320×320×64 feature map output from the first detection stage in the third concat connection module to obtain a 320×320×960 feature map. This feature map is then used by an attention-weighted c3k2 module with EMA to apply attention to the fused feature map. The 320×320×960 feature map output by this c3k2 module with EMA is primarily used for small target detection, specifically the detection of immature roses.

[0031] The 320×320×960 feature map output by the c3k2 module with EMA is input into the FasterNet-Pconv convolution module to adjust the channels. After being concatenated by the fourth Concat connection module, it is input into the c3k2 module. The 160×160×480 feature map obtained after processing by the c3k2 module is mainly used for mid-target detection, namely the detection of semi-ripe roses.

[0032] Meanwhile, the 160×160×480 feature map output by the c3k2 module is input into the EMA module for attention weighting. The feature map processed by the EMA module is then input into the FasterNet-Pconv convolution module for channel adjustment. In the fifth Concat connection module, it is concatenated and fused with the 40×40×512 feature map output from the fourth detection stage. The fused feature map is then passed through the c3k2 module of the EMA module for attention weighting, and its output 80×80×240 feature map is primarily used for large object detection, specifically the detection of mature roses.

[0033] The detection head in this application includes a small target detection head, a medium target detection head, and a large target detection head. The small target detection head is mainly used to detect immature roses, the medium target detection head is mainly used to detect semi-mature roses, and the large target detection head is mainly used to detect mature roses.

[0034] The three multi-scale feature maps output by the neck network are input into the corresponding target detection head. The detection head can detect the maturity of roses and predict their positions. The final detection results of the detection head include: the bounding box [x1, y1, x2, y2] of each rose, maturity label (0 / 1), and confidence score.

[0035] The third step is to train the model training set using the improved YOLOv11n-FE model to obtain the trained rose maturity detection model.

[0036] The fourth step is to input the image of the rose to be detected into the rose maturity detection model trained in the third step to detect the maturity of the target rose and obtain the final detection result.

[0037] The detection results obtained by this method include the bounding box, bounding box coordinates, maturity label, and confidence score for each rose in the rose image. All of these detection results are ultimately reflected in the rose image.

[0038] To verify the effectiveness of the improved YOLOv11n-FE model proposed in this application, the model was trained using the SGD optimizer on the rose maturity dataset constructed in this application, with the training epochs set to 300.

[0039] Recall (R), mean average precision (mAP), average precision (AP), precision (P), and model parameters are used as evaluation metrics. The formulas for R, mAP, AP, and P are as follows: TP represents the number of correctly identified positive samples, FN represents the number of incorrectly identified positive samples, N represents the total number of classes, and i represents the index variable for summation. The average precision for each category is: , , , , The improved YOLOv11n-FE model achieves optimizations in both mAP and parameter count. Compared to the existing YOLOv11n model, mAP is improved by 2.1%, and the number of parameters is reduced by 1.3M. Furthermore, a comparison with the YOLOv11n-FasterNet model shows that replacing the c3k2 module in the backbone network with the c3k2-FasterNet module reduces the number of model parameters by 1.4M without affecting mAP. This indicates that the introduction of the c3k2-FasterNet module effectively reduces the model's computational cost.

[0040] To verify the effectiveness of the EMA attention mechanism, comparative experiments were conducted using CA and SE attention mechanisms. The results show that, compared to the EUCB and YOLOv11n-EMA models, the YOLOv11n-FE model incorporating EMA exhibits a significant advantage in the rose maturity recognition task, with mAP improvements of 0.11% and 0.55%, respectively, while reducing the number of parameters by 0.7M and 0.2M. This result fully demonstrates that the introduction of the EMA attention mechanism improves the model's performance in rose maturity recognition. Table 1 shows a comparison of the precision and recall of the model used in this application with existing models.

[0041] Table 1. Comparison of precision and recall between the model used in this application and existing models.

[0042] While maintaining consistent experimental conditions, the YOLOv11n-FE model was compared with classic object detection algorithms YOLOv5n and SSD (VGG), as well as improved backbone models such as YOLOv8n-efficientVit and YOLOv9n-tiny. The results show that on the rose maturity recognition dataset, the improved YOLOv11n-FE model outperforms the other comparison models. Although the YOLOv9n-Snake model has a significant improvement in mAP compared to YOLOv11n, its increased parameter count leads to insufficient adaptability to devices with limited memory and low performance. Overall, the YOLOv11n-FE model, with its higher mAP and lower parameter count, is more suitable for deployment on low-performance devices for rose maturity recognition.

[0043] This application also includes a rose-picking robot. For example... Figures 2 to 4 As shown, the robot includes a body 2, with a collection basket 3 connected to the rear end of the body 2. The top of the collection basket 3 is open. Tracked walking mechanisms 6 are connected to the bottom of both sides of the body 2 along its direction of movement. The tracked walking mechanisms provide support for the entire robot and enable it to move.

[0044] The top surface of the vehicle body 2 is equipped with a robotic arm telescopic mechanism 7, a flexible rose holder mechanism 4, and a cutting and clamping mechanism 5. The free end of the robotic arm telescopic mechanism 7 is connected to the flexible rose holder mechanism 4, the cutting and clamping recognition camera 14, and the cutting and clamping mechanism 5. After the cutting and clamping mechanism 5 cuts the rose stems, the cut roses remain on the flexible rose holder mechanism 4. The robotic arm telescopic mechanism 7 then moves the flexible rose holder mechanism 4 to the top of the collection basket. The robotic arm telescopic mechanism 7 rotates, causing the cut and picked roses to fall into the collection basket 3.

[0045] The robotic arm telescopic mechanism 7 includes a rotary joint and a telescopic part 701. The telescopic part 701 includes a fixed branch pipe and a telescopic branch pipe. One end of the fixed branch pipe is connected to the top surface of the vehicle body through the rotary joint, and the telescopic branch pipe is slidably sleeved on the outside of the other end of the fixed branch pipe. One end of the telescopic branch pipe is slidably connected to the fixed branch pipe, and the other end of the telescopic branch pipe is connected to a flexible rose flower holder mechanism 4 and a cutting and clamping mechanism 5.

[0046] In this embodiment, both the rotary joint and the telescopic part 701 are driven by a high-torque motor 15. There are five rotary joints, so the robotic arm in this embodiment has five rotational degrees of freedom and one translational degree of freedom. The robotic arm has a large range of motion and can pick roses located in different positions.

[0047] In this embodiment, the fixed branch pipe and the telescopic branch pipe are slidably connected by a gear and rack assembly. The rack in the gear and rack assembly is fixed to the inner wall of the telescopic branch pipe. During the rotation of the gear driven by the high-torque motor 15, the meshing between the gear and rack converts the rotation of the gear into the linear movement of the rack, thereby causing the rack to drive the telescopic branch pipe to reciprocate along a straight line, thus realizing the relative sliding between the fixed branch pipe and the telescopic branch pipe in a straight line. The telescopic part enables the translational movement of the robotic arm, expanding the working space of the robotic arm during rose picking. At the same time, the gear and rack assembly has a relatively fast movement speed, which can reduce the impact on the working efficiency of the robotic arm.

[0048] The cutting and clamping mechanism includes a first connecting frame 501, a second connecting frame 502, and a cutting part. One end of the first connecting frame 501 is connected to the telescopic branch pipe of the robotic arm telescopic mechanism, and the other end of the first connecting frame 501 is rotatably connected to the second connecting frame 502. The other end of the second connecting frame 502 is fixed with the cutting part.

[0049] The cutting section includes a support frame 503, two opposing clamps, and two opposing cutters, all of which are mounted on the support frame 508. Figure 5As shown, the two clamps include a first clamp 504 and a second clamp 506. The first clamp 504 and the second clamp 506 are symmetrically arranged. The opposing surfaces of the first clamp 504 and the second clamp 506 are respectively provided with arc-shaped elastic clamping surfaces. When the first clamp 504 and the second clamp 506 are clamped together, the flower stem can be fixed without damaging the epidermis of the flower stem.

[0050] The two cutting devices include a first cutting device 505 and a second cutting device 507. The first cutting device 505 is fixedly connected to a first clamping device 504, and the second cutting device 507 is fixedly connected to a second clamping device 506. Single-edged cutting blades are respectively provided on the opposite surfaces of the first cutting device 505 and the second cutting device 507. While the first clamping device 504 and the second clamping device 506 clamp the flower stem, the cutting blades of the first cutting device 505 and the second cutting device 507 cut the flower stem.

[0051] In this embodiment, a gear set is used to open and close the two clamps and the two cutters. The ends of the first cutter and the first clamp, located on one side, are coaxial with the first gear. During rotation of the first gear, the first cutter and the first clamp rotate. The ends of the second cutter and the second clamp are coaxial with the second gear. During rotation of the second gear, the second cutter and the second clamp rotate. The first gear and the second gear mesh with each other. Therefore, when an external drive unit drives the first gear to rotate, the meshing between the first and second gears causes them to rotate in opposite directions, thereby causing the first cutter and the first clamp, and the second cutter and the second clamp, to rotate in opposite directions, thus realizing the opening and closing actions between the two cutters and the two clamps.

[0052] The flexible rose support mechanism 4 is fixedly connected to the cutting and clamping mechanism 5. The flexible rose support mechanism 4 includes a connecting plate 401 and support plates 402. One end of the connecting plate 401 is fixedly connected to the support plate 508 of the cutting and clamping mechanism, and the other end of the connecting plate 401 is fixed with two support plates 402. The two support plates 402 are symmetrically arranged with a gap between them. By reasonably setting the gap between the two support plates 402, the rose's flower support can be precisely held between the two support plates, thus supporting the rose. In this embodiment, the support plates are made of flexible silicone or low-hardness elastic polyurethane material, which can conform to the rose's flower support and avoid rigid damage.

[0053] A cutting and clamping recognition camera 14 is installed at the connection between the flexible rose holder mechanism and the cutting and clamping mechanism. The cutting and clamping recognition camera 14 is close to the cutting and clamping mechanism 5, and can accurately align with the flower stem to capture images of the flower stem to assist in the cutting operation of the flower stem and ensure the accuracy of image acquisition.

[0054] The top of the vehicle is also equipped with a supplementary light 1, a global camera 9, and a control mechanism 8. The tracked walking mechanism 6, the robotic arm telescopic mechanism 7, the flexible rose holder mechanism 4, the cutting and clamping mechanism 5, and the cutting and clamping recognition camera 14 are all connected to the control mechanism 8. The global camera 9 identifies and locates harvestable rose plants.

[0055] A storage layer is also provided below the robot body, where the camera navigation supplement light 13, power supply battery 12, and camera navigation camera 11 are all located. The power supply battery 12 provides power for the entire robot's operation. The camera navigation camera 11 records the robot's movement route, and the control mechanism 8 guides the robot's movement route based on the recorded movement route.

[0056] The tracked walking device 6 includes a track and a track drive mechanism. In this embodiment, the track drive mechanism includes a track drive gear 601. An internal rack is fixed on the inner surface of the track. The track drive gear 601 meshes with the internal rack. During the rotation of the gear, the internal rack is driven to move, thereby realizing the rotation of the track and the walking of the robot.

[0057] The two tracked walking devices 6 on both sides are connected to track drive motors 10, which drive the movement of each tracked walking device 6. Through the differential rotation of the two tracks, the robot can achieve movements such as forward and backward movement in a straight line, rotation in place, circular rotation, and braking. In actual use, the tracked walking devices 6 can automatically move forward, backward, turn, and rotate according to pre-set positioning markers.

[0058] This application also discloses a method for harvesting roses using the aforementioned rose-harvesting robot, the flowchart of which is shown below. Figure 6 As shown, it includes the following steps.

[0059] In the first step, the tracked walking device 6 drives the rose-picking robot to move between the rows of rose bushes, and the global camera 9 identifies the rose bushes. When a rose bush is detected, the global camera 9 and the cropping and clamping recognition camera 14 collect scene images of the rose bush and send the collected scene images to the control mechanism.

[0060] When there is no target to be picked in the field of view, the robot's global camera 9 and the cutting and clamping recognition camera 14 will rotate their angles to change the field of view and identify other rose plants to be picked within the working range.

[0061] The second step involves the YOLOv11n-FE model identifying roses on the rose plant based on the collected images, locating the roses, and determining their maturity. The YOLOv11n-FE model outputs the rose maturity recognition results, including the bounding box of each rose, the maturity level of the rose, and the confidence score.

[0062] The third step involves the control mechanism performing corresponding actions based on the rose maturity identification results output by the YOLOv11n-FE model.

[0063] When a mature rose is identified based on the rose maturity recognition results output by the YOLOv11n-FE model, the control mechanism will control the extension mechanism 7 of the robotic arm to move according to the position of the mature rose and complete the cutting action of the mature rose.

[0064] Specifically, the control mechanism sends a command to the robotic arm extension mechanism 7, which then moves towards the gripping point of the mature rose plant. If the extension and retraction of the robotic arm extension mechanism 7 alone is insufficient to move the cutting and gripping mechanism to the gripping point, the overall position of the robot needs to be adjusted appropriately.

[0065] After the position of the robotic arm telescopic mechanism 7 is adjusted, the flexible rose calyx mechanism 4 supports the lower part of the calyx of the mature rose, which plays a supporting role for the mature rose to be cut. The first gripper 504 and the second gripper 506 grab and hold the rose stem to fix it. At the same time, the first cutter 505 and the second cutter 507 perform the cutting action of the rose stem.

[0066] The pruned rose bushes, held by the first gripper 504 and the second gripper 506, rotate and move above the collection basket 3 along with the extension mechanism 7 of the robotic arm. When the rose bushes are directly above the collection basket 3, the first gripper 504 and the second gripper 506 activate, both grippers opening outwards. At this point, the gripping force on the rose bushes disappears, and the mature roses fall freely into the collection basket 3 under their own weight, completing the harvesting of one mature rose.

[0067] If there are other mature roses in the scene image obtained by the global camera 9 and the cropping and clamping recognition camera 14, repeat the operation of step three until all mature roses in the scene image have been picked.

[0068] When there is no target to pick in the field of view, the robot's global camera 9 and the cutting and clamping recognition camera 14 will rotate their angles to change the field of view and identify other rose plants to be picked within the working range until all the mature roses within the working range have been picked. Then the robot will continue to move along the ridges and enter the next working position.

[0069] The rose maturity identification method, rose picking robot, and rose picking method provided by this invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from the principles of this invention, and these improvements and modifications also fall within the protection scope of the claims of this invention. The above description of the disclosed embodiments enables those skilled in the art to implement or use this invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of this invention. Therefore, this invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying the maturity of roses, characterized in that, Includes the following steps: S1. Obtain images of roses to form a dataset; S2. Based on the improved YOLOv11n-FE model, the dataset is trained to obtain the trained rose maturity detection model; S3. Input the rose image to be detected into the rose maturity detection model obtained after training in step S2, perform target rose maturity detection, and obtain the final detection result.

2. The rose maturity identification method according to claim 1, characterized in that, The improved YOLOv11n-FE model includes a backbone network, a neck network, and a detection head.

3. The method for identifying rose maturity according to claim 2, characterized in that, The backbone network consists of four detection phases, each of which includes a FasterNet-Pconv convolutional module and several stacked c3k2-FsterNet modules; The first detection stage includes a FasterNet-Pconv convolutional module and three stacked c3k2-FsterNet modules: the input image for the first detection stage is a 640×640×3 image, which is passed through the 3×3 dimensionality reduction convolution of the FasterNet-Pconv convolutional module and the three stacked c3k2-FasterNet modules, and outputs a feature map with a size of 320×320×64. The second detection stage includes a FasterNet-Pconv convolutional module and six stacked c3k2-FsterNet modules: the input feature map of size 320×320×64 is passed through the 3×3 dimensionality reduction convolution of the FasterNet-Pconv convolutional module and the six stacked c3k2-FasterNet modules, and the output feature map of size 160×160×128 is passed. The third detection stage includes a FasterNet-Pconv convolutional module and six stacked c3k2-FsterNet modules: the input feature map of size 160×160×128 is passed through the 3×3 dimensionality reduction convolution of the FasterNet-Pconv convolutional module and the six stacked c3k2-FasterNet modules, and the output feature map of size 80×80×256 is passed. The fourth detection stage includes a FasterNet-Pconv convolutional module, three stacked c3k2-FsterNet modules, an SPPF module, and a C2PSA module. The input feature map of size 80×80×256 is passed through the 3×3 dimensionality reduction convolution of the FasterNet-Pconv convolutional module and the three stacked c3k2-FasterNet modules to output a feature map of size 40×40×512.

4. The method for identifying rose maturity according to claim 2, characterized in that, In the neck network, First, the 40×40×512 feature map output from the fourth detection stage is upsampled to 80×80 and then concatenated with the 80×80×256 feature map output from the third detection stage in the first concat connection module to obtain an 80×80×768 feature map. This feature map is then passed through the c3k2 module with EMA to perform attention weighting on the fused feature map to highlight maturity features. The processed 80×80×768 feature map is upsampled to 160×160, and then spliced ​​and fused with the 160×160×128 feature map output from the second detection stage in the second concat connection module to obtain a 160×160×896 feature map. This feature map is then used by the c3k2 module with EMA to perform attention weighting on the fused feature map to highlight maturity features. The processed 160×160×896 feature map is upsampled to 320×320, and then spliced ​​and fused with the 320×320×64 feature map output from the first detection stage in the third concat connection module to obtain a 320×320×960 feature map. This feature map is then used by the c3k2 module with EMA to perform attention weighting on the fused feature map, and the output 320×320×960 feature map is mainly used for small target detection, realizing the detection of immature roses; The output 320×320×960 feature map is input into the FasterNet-Pconv convolution module to adjust the channels, and then concatenated by the fourth Concat connection module before being input into the c3k2 module. The 160×160×480 feature map obtained after processing by the c3k2 module is mainly used for mid-target detection, realizing the detection of semi-ripe roses. The 160×160×480 feature map is input into the EMA module. After attention weighting by the EMA module, the feature map is input into the FasterNet-Pconv convolution module to adjust the channels. In the fifth Concat connection module, it is concatenated and fused with the 40×40×512 feature map output from the fourth detection stage. The fused feature map is then passed through the c3k2 module of the EMA module for attention weighting. The output 80×80×240 feature map is mainly used for large object detection, realizing the detection of mature roses.

5. A rose-picking robot, comprising a body, characterized in that, A collection basket is provided at the rear of the vehicle; The top of the vehicle body is equipped with a robotic arm telescopic mechanism. One end of the robotic arm telescopic mechanism is rotatably connected to the vehicle body, and the other end of the robotic arm telescopic mechanism is equipped with a flexible rose flower holder mechanism and a cutting and clamping mechanism. It also includes a control mechanism, with the robotic arm telescopic mechanism, the flexible rose petal holder mechanism, and the cutting and clamping mechanism all connected to the control mechanism.

6. The rose-picking robot according to claim 5, characterized in that, The robotic arm telescopic mechanism includes: Rotational joint; The telescopic part includes a fixed branch pipe and a telescopic branch pipe. One end of the fixed branch pipe is connected to the top of the vehicle body via a rotating joint. The other end of the fixed branch pipe is slidably connected to the end of the telescopic branch pipe. The other end of the telescopic branch pipe is connected to the cutting clamping mechanism.

7. The rose-picking robot according to claim 5, characterized in that, The cutting and clamping mechanism includes: The first connecting frame has one end fixedly connected to the free end of the robotic arm telescopic mechanism, and the other end rotatably connected to the second connecting frame. The second connecting frame is fixed with a support plate, and the support plate is fixedly connected to the cutting part and the clamping part. The clamping part includes a first clamp and a second clamp, which rotate in opposite directions, and both the first clamp and the second clamp are provided with arc-shaped elastic clamping surfaces on their opposing surfaces. The cutting section includes a first cutter and a second cutter. The first cutter is fixedly connected to a first clamp, and the second cutter is fixedly connected to a second clamp. Cutting blades are provided on the opposite surfaces of the first cutter and the second cutter.

8. The rose-picking robot according to claim 7, characterized in that, The flexible rose petal support mechanism includes: A connecting plate is fixedly connected to a support plate, and a cutting clamping recognition camera is provided at the fixed connection between the connecting plate and the support plate; The free end of the connecting plate has two symmetrically arranged support plates with a gap between them, which support the rose holder.

9. A method for harvesting roses using the rose-harvesting robot described in any one of claims 5-8, characterized in that, Includes the following steps: A1. The rose-picking robot moves between the rows of rose bushes and collects images of the rose bushes until it collects an image that has the characteristics of the target to be picked. A2. The improved YOLOv11n-FE model identifies roses on rose plants based on the collected images and outputs the rose maturity recognition results, including the bounding box of each rose, the maturity of the rose, and the confidence score. A3. Based on the rose maturity recognition results from the improved YOLOv11n-FE model, the identified mature roses are harvested: the robotic arm telescopic mechanism moves to support the flexible rose receptacle mechanism on the receptacle of the mature rose, the cutting and clamping mechanism completes the clamping, fixing and cutting of the mature rose stem, and collects the cut mature roses. A4. Once there are no more roses to pick in the field of vision, continue searching for rose plants and repeat steps A1 to A3 until all mature roses within the work area have been picked.

Images