A non-destructive measurement method for weight of lily bulbus of liliun davidii

Abstract

This invention relates to a non-destructive measurement method for the weight of Lanzhou lilies, comprising the following steps: S1 establishing a backbone network composed of a series of convolutional modules, C3k2 modules, and C2PSA modules at the ends; S2 establishing a small target-oriented dual-stream fusion pyramid; S3 establishing a multi-scale occlusion-focusing attention detection head; S4 establishing a YOLO-lily network architecture including a backbone network, a small target-oriented dual-stream fusion pyramid as the neck network, and a multi-scale occlusion-focusing attention detection head as the head network; S5 using the YOLO-lily network to perform instance segmentation on lily bulb images to obtain bulb mask images; S6 defining 22 bulb shape features based on the mask images and calculating the value of each feature; S7 using variance inflation factor evaluation to retain three key features: the volume of the circumscribed ellipsoid, the width of the circumscribed rectangle, and the length of the major axis of the circumscribed ellipse; S8 inputting the key features into a BPNN model to obtain the weight of the lily bulb. This invention achieves efficient and accurate estimation.

Description

Technical Field

[0001] This invention relates to the field of agricultural technology, and in particular to a non-destructive method for measuring the weight of Lanzhou lilies. Background Technology

[0002] Bulb weight is a core indicator determining the survival rate of Lanzhou lilies and is crucial for ensuring high yield and quality. However, current bulb weight assessment still mainly relies on manual weighing or experience-based estimation, which is time-consuming, labor-intensive, and difficult to adapt to large-scale production.

[0003] Existing research has demonstrated that image processing or geometric models can estimate phenotypic parameters of crop seeds. However, Lanzhou lily bulbs are irregular in shape and vary greatly in size, often covered with soil, and their density and moisture content fluctuate at different growth stages. Existing methods lack robustness to complex phenotypes, easily leading to excessive weight estimation errors. Moreover, in field harvesting scenarios, lily bulbs often suffer from stacking and occlusion, uneven lighting, and cluttered backgrounds, making it difficult for existing image processing methods to accurately extract bulb features. Furthermore, bulbs may exhibit different scales in images, especially smaller bulbs, making feature extraction even more challenging and further exacerbating the inaccuracy of weight estimation. These technical bottlenecks prevent existing image-driven weight estimation methods from being effectively applied to large-scale detection scenarios of Lanzhou lily bulbs, failing to meet the needs of the entire industry chain for precise bulb weight control. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a non-destructive measurement method for the efficient and accurate estimation of the weight of Lanzhou lily.

[0005] To solve the above problems, the present invention provides a non-destructive measurement method for the weight of Lanzhou lilies, comprising the following steps: S1 establishes a backbone network consisting of a series of convolutional modules, C3k2 modules, and C2PSA modules at the ends; S2 establishes a small-goal-oriented dual-stream fusion pyramid: Using the multi-scale features of P2:160×160×C, P3:80×80×C, P4:40×40×C, and P5:20×20×C in the backbone network as input, the features are processed by the convolution module and the C3k2 module to complete the lateral feature enhancement, and multi-scale feature fusion is performed along the bottom-up and top-down paths respectively. S3 establishes a multi-scale occlusion-focused attention detection head: A three-way parallel DCovN feature extraction structure is adopted; each DCovN unit is configured with a multi-size convolutional kernel group to form a hierarchically distributed receptive field; S4 establishes a YOLO-lily network architecture that includes a backbone network, a small-target-oriented dual-stream fusion pyramid as the neck network, and a multi-scale occlusion-focused attention detection head as the head network. S5 uses the YOLO-lily network to perform instance segmentation on lily bulb images and obtain bulb mask images; S6 defines 22 spherical shape features based on mask images and calculates the feature values ​​using corresponding calculation formulas. S7 uses the variance inflation factor to assess multicollinearity, eliminates redundant features, and retains three key features: the volume of the circumscribed ellipsoid, the width of the circumscribed rectangle, and the length of the major axis of the circumscribed ellipse. S8 inputs key features into the BPNN model to obtain the weight of the lily bulb.

[0006] In step S2, the bottom-up path refers to the model performing bottom-up downsampling and feature fusion operations from P2 to P5: the C3k2 module downsamples high-resolution features through convolution with a stride of 2, then fuses them with adjacent high-level semantic features through a concatenation operation, and completes feature dimensionality reduction and integration using 1×1 convolution.

[0007] In step S2, the top-down path refers to the model performing top-to-bottom upsampling and feature fusion operations from P5 to P2: the C3k2 module first enhances the high-level semantic features, then aligns them with the low-level features of adjacent layers through bilinear upsampling, and completes feature fusion by adding elements one by one.

[0008] In step S3, the multi-size convolutional kernel group consists of a small-size convolutional kernel focused on mining the fine local features of the target, a medium-size convolutional kernel used to capture the local relationship between the target and the occluder, and a large-size convolutional kernel that extracts global contextual information and uses the global semantic association of the scene to achieve feature completion and inference of the occluded area.

[0009] The workflow of the DCovN unit in step S3 is as follows: the input feature map is adaptively divided into several non-overlapping feature blocks and mapped to a high-dimensional feature space; features are extracted by stacked deep convolutional modules and pointwise convolutional modules; global average pooling is performed on the output feature map and the original input feature map respectively and then fused; channel attention weights are calculated through a fully connected layer; and the weight differences are amplified by exponential operation and applied to the feature map for adjustment, so as to achieve accurate focusing of the occluded seed ball features.

[0010] In step S6, the 22 seed bulb shape characteristics refer to the seed bulb's projected area ( A mm^2 ), bulb outline circumference ( P mm ), shape complexity ( SC), length of the major axis of the circumscribed ellipse ( MA ), the length of the minor axis of the circumscribed ellipse ( mA ), circumspherical volume ( V i ), the diameter of the circumscribed ellipse ( D ), the extensibility of the circumscribed ellipse ( E ), convex hull area ( A ch_mm^2 ), convex hull perimeter ( P ch_mm ), ratio of bulb area to convex hull area ( CR ), convex hull roundness ( C ch ), area of ​​the circumscribed rectangle ( A rect_mm^2 ), aspect ratio of the circumscribed rectangle ( AR rect ), outer rectangle width ( W rect_mm ), height of the bounding rectangle ( H rect_mm ), minimum circumcircle diameter ( D circle ), circumscribed cuboid volume ( V box ), height to area ratio ( HAR Diameter based on area () D area ), the maximum diameter of the circumscribed rectangle ( D span ), Area and Complexity Interaction Features ( ACI ).

[0011] Compared with the prior art, the present invention has the following advantages: 1. This invention, through the collaborative innovation of "small target-oriented dual-stream fusion pyramid" and "multi-scale occlusion-focusing attention detection head", combined with the YOLO-Lily exclusive network architecture and accurate weight prediction model, fundamentally solves the three core problems faced in the weight estimation of Lanzhou lily bulbs: "poor adaptability to complex scenarios", "low accuracy of small target extraction" and "large weight estimation error", and achieves efficient and accurate estimation of lily bulb weight in large-scale production scenarios.

[0012] 2. Technological paradigm breakthrough: From general image processing to end-to-end precise estimation of seed bulbs.

[0013] Existing technologies: Traditional bulb weight estimation relies on manual weighing or experience-based estimation, which is inefficient; existing image-driven weight estimation methods are mostly general image processing or geometric modeling solutions, which are not designed for the specific characteristics of Lanzhou lily bulbs, such as "irregular shape, mud on the surface, and large size difference". They also do not fully consider the complex environment in the field harvesting scene, such as stacking and shading, uneven lighting, and cluttered background. They are not good at extracting features of small-sized bulbs, resulting in poor robustness, large weight estimation errors, and inability to meet the needs of large-scale production.

[0014] This invention completely abandons the generalized processing approach and constructs a dedicated end-to-end estimation system for Lanzhou lily bulbs. Using the YOLO-lily network as its core, it adapts customized neck and head modules to bulb characteristics, combining them with a dedicated shape feature definition and weight prediction model to achieve integrated processing from image acquisition and accurate bulb extraction to weight estimation. This accurately matches the complex phenotype of Lanzhou lily bulbs and the complex field environment, significantly improving estimation accuracy and robustness.

[0015] 3. Breakthrough in small target detection: From single-scale feature fusion to dual-stream guided multi-scale enhancement fusion.

[0016] Existing technology: Feature fusion in existing target detection models often adopts a single-direction or single-scale fusion method. The high-level feature map output by the backbone network has a weak perception ability for small targets such as lily bulbs with extremely low pixel ratio and diverging roots. It cannot effectively extract feature information of small bulbs, resulting in high rates of missed detection and false detection of small target bulbs, which in turn affects the accuracy of weight estimation.

[0017] This invention innovatively designs a small-target-oriented dual-stream fusion pyramid, introducing a high-resolution P2 layer. Using multi-scale features (P2-P5) as input, it employs a bottom-up and top-down bidirectional fusion path, combined with the lateral feature enhancement and differentiated fusion strategy of the C3k2 module, to fully integrate the shallow detail features and deep semantic features of seed bulbs of different scales. In particular, it strengthens the feature expression of small-sized seed bulbs, significantly improving the extraction accuracy of small-target seed bulbs and solving the technical bottleneck of difficult feature extraction of small-scale seed bulbs.

[0018] 4. Breakthrough in occlusion scene adaptation: From conventional feature extraction to multi-scale occlusion-focused attention enhancement.

[0019] Existing technology: Current image processing methods do not have a dedicated feature extraction mechanism designed for occlusion scenarios such as stacked lily bulbs or mud on the surface. Conventional convolutional kernels are difficult to distinguish between bulb features and occlusion features (soil, vegetation, tools), and are easily affected by occlusion interference, resulting in incomplete bulb feature extraction and failure to achieve accurate identification and positioning of bulbs under occlusion conditions.

[0020] This invention innovatively proposes a multi-scale occlusion-focused attention detection head, employing a three-way parallel DCovN structure and multi-size convolutional kernel groups to construct a hierarchical receptive field architecture. This accurately captures local details of the seed ball, occlusion relationships, and global contextual information. By separating and enhancing the attention mechanism and weight redistribution strategy, it focuses on the seed ball features in occluded areas, suppresses interference from occluders, and significantly improves the model's adaptability to complex occlusion scenarios, thus solving the technical defect of insufficient seed ball extraction accuracy under occlusion conditions. Attached Figure Description

[0021] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0022] Figure 1 This invention relates to a small-target-oriented dual-stream fusion pyramid.

[0023] Figure 2 This invention relates to a multi-scale occlusion focusing attention detection head.

[0024] Figure 3 This is the YOLO-lily network architecture of the present invention.

[0025] Figure 4 The key metrics changes during the training process of the YOLO-lily network of this invention are shown below. Where: a is the segmentation loss; b is the bounding box regression loss; c is the average accuracy when the IoU threshold is 0.5; and d is the average accuracy when the IoU threshold is between 0.5 and 0.95.

[0026] Figure 5 This invention provides a comparison of YOLO-lily with other instance segmentation models.

[0027] Figure 6 This is a scatter plot of the predicted and true values ​​of lily bulb weight according to the present invention.

[0028] Figure 7 This is a diagram showing the results of non-destructive measurement of the weight of Lanzhou lily bulbs according to the present invention. Detailed Implementation

[0029] This invention constructs a multi-scale occlusion-focused attention detection head and a small-target-oriented dual-stream fusion pyramid, building upon Yolo V11 to construct Yolo-lily, achieving high-precision extraction of lily bulbs from digital images, and subsequently establishing a bulb weight estimation model. Specifically: the multi-scale occlusion-focused attention detection head can effectively focus on bulb features in areas covered by soil or stacked occlusion, enhancing the model's robustness to complex phenotypes and occlusion scenarios; the small-target-oriented dual-stream fusion pyramid can fully integrate bulb features of different scales, especially strengthening the feature representation of small-sized bulbs, improving the extraction accuracy of small-target bulbs.

[0030] A non-destructive method for measuring the weight of Lanzhou lilies includes the following steps: S1 establishes a backbone network consisting of a series of convolutional modules, C3k2 modules, and C2PSA modules at the ends.

[0031] The convolutional module is mainly used for feature extraction, while the C3k2 module is responsible for feature separation, convolution, and fusion operations. To improve the spatial awareness and global context modeling capabilities of the model in this invention, a C2PSA module is introduced at the end of the backbone network, thereby improving the model's accuracy in perceiving the geometric shape and complex spatial relationships of the lily bulb.

[0032] S2 establishes a small-target-oriented dual-stream fusion pyramid for precise extraction of lily bulbs, such as Figure 1 As shown: For lily bulbs with divergent rootlets and extremely low pixel ratios, the high-level features output by the backbone network are insufficient for effective detection. This invention uses multi-scale features from the backbone network (P2: 160×160×C, P3: 80×80×C, P4: 40×40×C, P5: 20×20×C) as input. These features are processed by convolutional and C3k2 modules to enhance lateral features. Multi-scale feature fusion is then performed along both bottom-up and top-down paths to achieve efficient integration of multi-scale lily bulb features, strengthening the representation of small target bulb features and thus improving the model's detection and segmentation capabilities for lily bulbs.

[0033] Among them, the bottom-up path refers to the model performing bottom-up downsampling and feature fusion operations from P2 to P5: the C3k2 module downsamples high-resolution features through convolution with a stride of 2, then fuses them with adjacent high-level semantic features through a concatenation operation, and completes feature dimensionality reduction and integration using 1×1 convolution.

[0034] The top-down path refers to the model performing top-to-bottom upsampling and feature fusion operations from P5 to P2: the C3k2 module first enhances the high-level semantic features, then aligns them with the low-level features of adjacent layers through bilinear upsampling, and completes feature fusion by adding elements one by one.

[0035] To verify the effectiveness of the small target-oriented dual-stream fusion pyramid (ST-DSFP), it was compared with feature fusion modules such as GFPN, GOLDYOLO, HSFPN, SDI and BiFPN. The experimental results are shown in Table 1.

[0036] Table 1 Model Comparison Experiment As shown in Table 1, the ST-DSFP module requires only 1.83M parameters and 5.8G of model computation (FLOPs). While maintaining low computational overhead, it achieves a detection accuracy of 78.8% and a mAP@0.5 of 68.2%, both of which are superior to other modules. In comparison, GFPN also achieves a detection accuracy of 78.3%, but its parameter count and computational cost are 100% and 41% higher than ST-DSFP, respectively. GOLDYOLO has the lowest performance. Although BiFPN is similar to ST-DSFP in terms of parameter count and computational cost, its detection accuracy and mAP@0.5 are still slightly lower than DSFP. Furthermore, HSFPN and SDI each have their advantages in low parameter count and mAP@0.5, but ST-DSFP balances model complexity and detection capability, improving accuracy and mAP@0.5 by 1.3% and 2.0%, respectively, with a similar parameter count to HSFPN. This demonstrates that ST-DSFP can more effectively balance image detail features and deep semantic features during the bidirectional feature flow process, thereby significantly improving the perception capability of small-sized, fast-moving targets.

[0037] S3 establishes a multi-scale occlusion-focused attention detection head for occlusion-focused lily bulb extraction, such as Figure 2 As shown: To address the shortcomings of existing technologies, such as insufficient accuracy in detecting occluded targets and poor adaptability to complex occlusion scenarios, this invention employs targeted structural design to redistribute the weights of feature maps, effectively enhancing the segmentation and detection performance of lily bulbs under occlusion and significantly improving the model's ability to identify and locate occluded targets.

[0038] Specifically, the detection head employs a three-way parallel DCovN feature extraction structure to construct a multi-scale feature perception system, enabling omnidirectional feature capture of occluded targets at different scales. Each DCovN unit is configured with a multi-size convolutional kernel group, forming a hierarchically distributed receptive field through differentiated design of the convolutional kernel scale.

[0039] The multi-size convolutional kernel group consists of small-sized convolutional kernels focused on mining fine local features of the target, medium-sized convolutional kernels used to capture the local relationships between the target and occluders, and large-sized convolutional kernels that extract global contextual information and achieve feature completion and inference of occluded regions by leveraging global semantic relationships of the scene. Small-sized convolutional kernels ensure accurate perception of residual details after occlusion; medium-sized convolutional kernels provide feature support for defining the boundaries of occluded regions; and large-sized convolutional kernels effectively overcome the occlusion interference of occluders on target features.

[0040] The workflow of the DCovN unit is as follows: The input feature map is adaptively divided into several non-overlapping feature blocks and mapped to a high-dimensional feature space. Features are extracted through stacked deep convolutional modules and pointwise convolutional modules. Global average pooling is performed on the output feature map and the original input feature map, and then fused. Channel attention weights are calculated through a fully connected layer, and the weight differences are amplified by exponential operations before being applied to the feature map for adjustment, achieving precise focusing of occluded seed features. The specific process is as follows: First, the input feature map is adaptively divided into several non-overlapping feature blocks, and transformed into a high-dimensional feature space through a feature mapping mechanism to enhance the dimensionality and representation of feature information. Then, a feature extraction chain is constructed by stacking deep convolutional and pointwise convolutional modules. The deep convolutional module independently performs convolution operations on the feature information of each channel, ensuring the integrity and independence of single-channel features, while the pointwise convolutional module focuses on cross-channel feature fusion, achieving complementarity and integration of feature information from different channels, and improving the comprehensive representation capability of features. Finally, global average pooling is performed on the output feature map of each DCovN unit and the original input feature map. The multiple sets of feature information obtained from pooling are fused and input into a fully connected layer. The attention weights of each channel are calculated through the nonlinear transformation of the fully connected layer. Then, exponential operations are used to amplify the differences in attention weights between different channels, strengthening the weight allocation for occluded target feature channels. Finally, the optimized attention weights are applied to the adaptive adjustment of the feature map, achieving precise focusing and significant enhancement of occluded target feature information, effectively suppressing background and occlusion interference.

[0041] To verify the effectiveness of the Multi-Scale Occlusion Focused Attention Detector (MOA-detect), a comparative experiment was conducted to comprehensively compare MOA-detect with various attention modules such as BAMBlock, CoordAtt, LSKA, LSKBlock, MLCA, MPCA, SEAttention, SimAM, and TripletAttention. The experimental results are shown in Table 2.

[0042] Table 2 Comparative Experiment of Attention Modules As shown in Table 2, MOA-detect achieves an accuracy of 79.5% and mAP@0.5 reaches 69.7%, significantly outperforming other attention modules.

[0043] S4 establishes a YOLO-lily network architecture that includes a backbone network, a small-target-oriented dual-stream fusion pyramid as the neck network, and a multi-scale occlusion-focused attention detection head as the head network (e.g., Figure 3 As shown in the figure, the three work together to achieve real-time and accurate instance segmentation of lily bulbs.

[0044] YOLO-lily is a real-time lily bulb instance segmentation model. The model consists of three parts: a backbone network, a neck network, and a head network, which are responsible for feature extraction, feature fusion, and target segmentation and localization, respectively.

[0045] Five hundred Lanzhou lily bulbs were selected to establish a YOLO-lily model according to the method of this invention, and the model was trained as follows: the Adaptive Moment Estimator (AdamW) was used, and the initial learning rate was set to 1×10⁻⁶. −4 The weight decay coefficient is 5×10 −4 The batch size is 16; during training, a cosine annealing learning rate scheduling strategy is adopted, and a total of 200 rounds are trained. In the first 10 rounds, a warm-up mechanism is used to gradually increase the learning rate to avoid gradient oscillation in the early stage of training.

[0046] The segmentation loss function is constructed by weighting and combining Dice loss and cross-entropy loss. L seg The weighting ratio is 0.7:0.3; the splitting loss is statistically analyzed on the validation set according to the round. L seg Bounding box regression loss L box The segmentation accuracy mAP50 and mean accuracy mAP50-95 were calculated, and training curves were plotted. The results are as follows: Figure 4 As shown. By Figure 4 (ab) shows that the segmentation loss value L seg The loss curve initially dropped rapidly from approximately 11.0, with a significant decrease within the first 15 training epochs, indicating effective optimization of the segmentation head parameters. Subsequently, the loss curve plateaued, stabilizing around 0.25 after 80 training epochs, suggesting that the mask prediction learning process had essentially reached a stable state. (Bounding box regression loss) L box The segmentation accuracy (mAP50) decreased from an initial value of 2.8 to around 0.5 over 100 training epochs. The segmentation accuracy mAP50 increased sharply from 0, reaching a critical inflection point of 0.85 at the 25th epoch, and eventually converged to 0.98 (see [link to training epoch]). Figure 4 c) The early saturation characteristic of this metric reflects the model's ability to quickly identify standard samples. The mean accuracy (mAP) of 50-95 increases rapidly at the 50th round, reaching 0.65, and eventually reaches 0.95 after continuous training (see [link]). Figure 4 (d) This indicates that under a high Intersection over Union (IoU) threshold, the model requires a longer optimization process to achieve accurate localization. The smooth curve trajectory shows that all metrics tend to stabilize after 100 training epochs, without any oscillations caused by overfitting.

[0047] S5 uses the YOLO-lily network to perform instance segmentation on lily bulb images and obtain bulb mask images.

[0048] Using 24 lily bulbs numbered 1, 2, 3, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 19, 20, 21, 22, 23, 24, 25, 26, 27, and 29 as the original images, the results were compared with mainstream segmentation models such as YOLO-lily, SOLOv2, Mask-RCNN, and Queryinst. The results are as follows: Figure 5 As shown in the figure, the SOLOv2 model failed to accurately reproduce the circular outline of the bulb, and the segmentation edges of the calibration markers showed severe distortion; the bulb mask generated by Mask-RCNN deviated from the actual shape of the bulb, resulting in insufficient morphological representation accuracy; the Queryinst model exhibited bulb omission issues in some test images. In contrast, the YOLO-lily model effectively avoided the above defects, demonstrating stable and accurate segmentation performance.

[0049] S6 defines 22 spherical shape features based on the mask image and calculates the feature values ​​using the corresponding calculation formulas.

[0050] Among them: 22 seed bulb shape characteristics refer to the seed bulb's projected area ( A mm^2 ), bulb outline circumference ( P mm ), shape complexity ( SC ), length of the major axis of the circumscribed ellipse ( MA ), the length of the minor axis of the circumscribed ellipse ( mA ), circumspherical volume ( V i ), the diameter of the circumscribed ellipse ( D ), the extensibility of the circumscribed ellipse ( E ), convex hull area ( A ch_mm^2 ), convex hull perimeter ( P ch_mm ), ratio of bulb area to convex hull area ( CR ), convex hull roundness ( C ch ), area of ​​the circumscribed rectangle ( A rect_mm^2 ), aspect ratio of the circumscribed rectangle ( AR rect ), outer rectangle width ( W rect_mm ), height of the bounding rectangle ( H rect_mm ), minimum circumcircle diameter ( Dcircle ), circumscribed cuboid volume ( V box ), height to area ratio ( HAR Diameter based on area () D area ), the maximum diameter of the circumscribed rectangle ( D span ), Area and Complexity Interaction Features ( ACI (as shown in Table 3).

[0051] Table 3. Characteristics of seed bulb shape S7 uses the variance expansion factor to assess multicollinearity, eliminates redundant features, and retains three key features: the volume of the circumscribed ellipsoid, the width of the circumscribed rectangle, and the length of the major axis of the circumscribed ellipse.

[0052] By inputting key features into the BPNN model, the S8 model can obtain the weight of lily bulbs, thus achieving accurate prediction of lily bulb weight.

[0053] Example 500 Lanzhou lily bulbs were selected and a YOLO-lily model was established according to the method of this invention. The shape parameters of the 500 bulbs were statistically analyzed, and the results are shown in Table 4.

[0054] Table 4 Shape Parameters Multicollinearity among features was assessed using the variance inflation factor. Redundant features were eliminated, retaining three key features: the volume of the circumscribed ellipsoid, the width of the circumscribed rectangle, and the length of the major axis of the circumscribed ellipse. Inputting these key features into a BPNN model yields the weight of the lily bulb, such as... Figures 6-7 As shown, its R² reaches 0.9675, MSE is 2.2424, and MAE is 0.8066.

[0055] In summary, by using a small-target-oriented dual-stream fusion pyramid and a multi-scale occlusion-focused attention detection head to improve YOLO V11 and construct a YOLO-lily-specific network architecture, the instance segmentation accuracy of Lanzhou lily bulbs in complex scenarios was effectively improved. Furthermore, through precise shape feature screening and a BPNN weight prediction model, efficient, lossless, and accurate estimation of bulb weight was finally achieved, with an R² of 0.9675. This provides reliable technical support for the large-scale grading and industrial application of Lanzhou lily bulbs.

Claims

1. A non-destructive method for measuring the weight of Lanzhou lilies, comprising the following steps: S1 establishes a backbone network consisting of a series of convolutional modules, C3k2 modules, and C2PSA modules at the ends; S2 establishes a small-goal-oriented dual-stream fusion pyramid: Using the multi-scale features of P2:160×160×C, P3:80×80×C, P4:40×40×C, and P5:20×20×C in the backbone network as input, the features are processed by the convolution module and the C3k2 module to complete the lateral feature enhancement, and multi-scale feature fusion is performed along the bottom-up and top-down paths respectively. S3 establishes a multi-scale occlusion-focused attention detection head: A three-way parallel DCovN feature extraction structure is adopted; each DCovN unit is configured with a multi-size convolutional kernel group to form a hierarchically distributed receptive field; S4 establishes a YOLO-lily network architecture that includes a backbone network, a small-target-oriented dual-stream fusion pyramid as the neck network, and a multi-scale occlusion-focused attention detection head as the head network. S5 uses the YOLO-lily network to perform instance segmentation on lily bulb images and obtain bulb mask images; S6 defines 22 spherical shape features based on mask images and calculates the feature values ​​using corresponding calculation formulas. S7 uses the variance inflation factor to assess multicollinearity, eliminates redundant features, and retains three key features: the volume of the circumscribed ellipsoid, the width of the circumscribed rectangle, and the length of the major axis of the circumscribed ellipse. S8 inputs key features into the BPNN model to obtain the weight of the lily bulb.

2. The non-destructive measurement method for the weight of Lanzhou lilies as described in claim 1, characterized in that: In step S2, the bottom-up path refers to the model performing bottom-up downsampling and feature fusion operations from P2 to P5: the C3k2 module downsamples high-resolution features through convolution with a stride of 2, then fuses them with adjacent high-level semantic features through a concatenation operation, and completes feature dimensionality reduction and integration using 1×1 convolution.

3. The non-destructive measurement method for the weight of Lanzhou lilies as described in claim 1, characterized in that: In step S2, the top-down path refers to the model performing top-to-bottom upsampling and feature fusion operations from P5 to P2: the C3k2 module first enhances the high-level semantic features, then aligns them with the low-level features of adjacent layers through bilinear upsampling, and completes feature fusion by adding elements one by one.

4. The non-destructive measurement method for the weight of Lanzhou lilies as described in claim 1, characterized in that: In step S3, the multi-size convolutional kernel group consists of a small-size convolutional kernel focused on mining the fine local features of the target, a medium-size convolutional kernel used to capture the local relationship between the target and the occluder, and a large-size convolutional kernel that extracts global contextual information and uses the global semantic association of the scene to achieve feature completion and inference of the occluded area.

5. The non-destructive measurement method for the weight of Lanzhou lilies as described in claim 1, characterized in that: The workflow of the DCovN unit in step S3 is as follows: the input feature map is adaptively divided into several non-overlapping feature blocks and mapped to a high-dimensional feature space; features are extracted by stacked deep convolutional modules and pointwise convolutional modules; global average pooling is performed on the output feature map and the original input feature map respectively and then fused; channel attention weights are calculated through a fully connected layer; and the weight differences are amplified by exponential operation and applied to the feature map for adjustment, so as to achieve accurate focusing of the occluded seed ball features.

6. The non-destructive measurement method for the weight of Lanzhou lilies as described in claim 1, characterized in that: The 22 seed sphere shape features in step S6 refer to the seed sphere's projected area, seed sphere's outline perimeter, shape complexity, length of the major axis of the circumscribed ellipse, length of the minor axis of the circumscribed ellipse, volume of the circumscribed ellipsoid, diameter fitted by the circumscribed ellipse, extensibility of the circumscribed ellipse, convex hull area, convex hull perimeter, ratio of seed sphere region to convex hull region, convex hull roundness, area of ​​the circumscribed rectangle, aspect ratio of the circumscribed rectangle, width of the circumscribed rectangle, height of the circumscribed rectangle, minimum diameter of the circumscribed circle, volume of the circumscribed cuboid, height to area ratio, diameter based on area, maximum diameter of the circumscribed rectangle, and interaction features between area and complexity.

Images