Distillation-aware structured channel pruning method and system for bulb plant visual recognition
Patent Information
- Application Number
- CN202611108018.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-24
- Publication Date
- 2026-10-13
AI Technical Summary
但单一压缩技术存在显著缺陷：当压缩率超过80%时，仅靠知识蒸馏易丢失教师网络中的细粒度判别特征，仅靠剪枝容易造成关键通道的错误移除，仅靠量化则会引入不可忽略的量化误差，导致球茎霉斑识别或成熟度判断精度急剧下降
[0042]（1）现有Network Slimming等结构化剪枝方法在压缩率超过80%时已出现明显的精度崩塌。本发明在压缩率超过90%时仍能保持对微小霉斑的识别能力；且通过注意力迁移初始化学生网络并联合知识蒸馏进行剪枝，使剪枝过程直接受教师网络表征能力的引导，即使在压缩率超过80%乃至达到90%以上的极端情况下，分类或检测精度下降仍可控制在2个百分点以内。
Smart Images
Figure CN122904981A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of lightweight technology of edge intelligence and deep learning models, specifically involving a distillation-sensory structured channel pruning method and system for visual recognition of bulbous plants; it is particularly suitable for performing fine-grained visual tasks such as bulb mold detection, bulb epidermal maturity judgment, leaf disease identification and obstacle avoidance in real time on resource-constrained edge computing devices. Background Technology
[0002] In 5G-connected smart bulb plant cultivation scenarios, vehicle-mounted edge devices need to run high-precision visual perception models in real time to complete fine-grained tasks such as bulb mold detection, bulb epidermal texture analysis, and leaf color gradient recognition. However, high-precision models such as ResNet and Vision Transformer (ViT) have large parameter scales and high computational complexity, making it difficult to meet the real-time inference requirement of less than 20 milliseconds / frame on edge devices with limited computing power, storage, and power consumption. Therefore, the models must be significantly compressed.
[0003] Existing model compression techniques mainly include knowledge distillation (such as the knowledge distillation method proposed by Hinton et al. in 2015), network pruning (such as Network Slimming and ThiNet methods, Liu et al., 2017), and weight quantization (such as the weight quantization method proposed by Jacob et al. in 2018). However, single compression techniques have significant drawbacks: when the compression rate exceeds 80%, knowledge distillation alone is prone to losing fine-grained discriminative features in the teacher network, pruning alone is prone to causing erroneous removal of key channels, and quantization alone will introduce non-negligible quantization errors, leading to a sharp decline in the accuracy of bulb mold spot recognition or maturity judgment.
[0004] On the other hand, while unstructured pruning can significantly reduce parameters, the resulting sparse matrix is difficult to achieve practical acceleration on general-purpose hardware and cannot effectively reduce inference latency on edge devices.
[0005] More importantly, general compression methods are guided only by global metrics such as classification accuracy, completely ignoring the specificities of visual tasks in bulbous plants. For example, mold spots with a diameter of only a few pixels on the bulb epidermis, the delicate epidermal texture and specific color gradients produced by changes in bulb maturity, and early signs of disease on leaves are easily regarded as redundant information and discarded during general pruning, resulting in severe degradation of the performance of compressed models in real-world horticultural scenarios.
[0006] Furthermore, existing pruning methods based on channel scaling factors (such as Network Slimming and its improvements) rely solely on global sparse regularization to guide channel selection, lacking protection mechanisms for features sensitive to specific visual tasks. Under extremely high compression rates, they are prone to erroneously pruning discriminative fine-grained channels.
[0007] Therefore, there is an urgent need for a structured model compression scheme that can retain the sensitive characteristics of bulbous plants even at extremely high compression rates. Summary of the Invention
[0008] The purpose of this invention is to propose a distillation-based perceptual structured channel pruning method and system for visual recognition of bulbous plants. By combining teacher network knowledge distillation, structured sparse learning, and bulbous plant feature preservation constraints, the accuracy of key visual tasks such as bulb mold spot recognition and maturity judgment is reduced by no more than 2 percentage points under conditions of significant reduction in parameter and computational load. Moreover, the compressed regular network structure can directly achieve significant inference acceleration on edge devices. Specifically, distillation perception refers to using the attention feature map of the teacher network as a supervision signal during the pruning process, guiding the student network to maintain the key feature representation ability learned by the teacher network during parameter compression through knowledge distillation loss.
[0009] To achieve the above objectives, the technical solution of the present invention is as follows:
[0010] A distillation-sensory structured channel pruning method for visual recognition of bulbous plants includes the following steps:
[0011] Step S1: Obtain the teacher network trained to convergence on the target task dataset (on a cloud GPU cluster), and construct an initial student network with a reduced number of channels; wherein, the number of output channels of each layer of the initial student network can be set to 40% to 60% of the number of channels of the corresponding layer of the teacher network, and further compression can be achieved through channel pruning, with a total compression rate exceeding 80%;
[0012] Step S2: Using the attention transfer mechanism, the spatial attention map corresponding to the intermediate layer feature map specified by the teacher network is used as the supervision target to initialize the parameters of the initial student network, so that the student network has a similar feature attention region to the teacher network at the beginning of training.
[0013] Step S3: Introduce a learnable channel scaling factor (the channel scaling factor is multiplicatively applied to the output feature map of the corresponding channel) for each convolutional layer of the student network, and apply a sparse regularization term during training to push a large number of scaling factors close to zero;
[0014] Step S4: Construct a joint loss function to perform sparse training on the student network containing the channel scaling factor; the joint loss function includes at least the task loss (to ensure the discriminative ability of the student network on the bulb plant visual task), the distillation loss for aligning the spatial attention maps corresponding to the feature maps of the teacher network and the student network in the specified intermediate layer, the sparse regularization loss acting on the channel scaling factor (to induce the network to produce structured sparsity), and the bulb plant feature preservation loss for maintaining the consistency of features between the teacher network and the student network in the three sensitive regions of the plant (bulb mold texture, bulb epidermis, and leaf color gradient);
[0015] Step S5: After sparse training is completed, based on the size of the channel scaling factor, all convolutional channels with scaling factors lower than the pruning threshold are pruned to obtain a compact and regularly numbered structured pruning model.
[0016] Step S6: Fine-tune the structured pruning model using the target task dataset to recover the slight accuracy perturbation caused by channel removal, and obtain a lightweight model that can be used for real-time inference on edge devices for bulb plant visual recognition.
[0017] Preferably, the bulb plant feature preservation loss includes: leaf color gradient preservation loss acting on the output of the first network stage of the teacher network and student network, bulb mold texture preservation loss acting on the output of the second network stage, and bulb epidermis preservation loss acting on the third network stage, forming a multi-level feature preservation mechanism from low to high level.
[0018] Preferably, the leaf color gradient preservation loss is constructed in the following manner:
[0019] Calculate the gradient magnitude of the input image, and take the region with a gradient magnitude greater than a preset threshold as the mask for the region with significant leaf color gradient;
[0020] The feature maps output by the teacher network and the student network in the first network stage are globally averaged along the channel direction to obtain single-channel feature maps.
[0021] Calculate the weighted average vector of the gradient magnitude of the single-channel feature map within the mask region, and use L2 distance to constrain the weighted average vectors of the teacher network and the student network to tend to be consistent.
[0022] Preferably, the loss of bulb mold texture retention is constructed in the following way:
[0023] Obtain the mask of the bulb mold patch region in the input image;
[0024] For the feature maps output from the second network stage of the teacher network and the student network, the features of the moldy area are extracted separately using the bulb moldy area mask.
[0025] The extracted mold patch region features were convolved using a Gabor filter bank to obtain multiple sets of texture response maps;
[0026] Calculate the Gram matrix of the texture response map within the mold patch region;
[0027] The difference between the Gram matrices of the teacher network and the student network is measured using the Frobenius norm, and is used as the bulb mold texture preservation loss.
[0028] Preferably, the texture preservation loss of the bulb mold spot is extracted by using a Gabor filter bank with 4 directions and 3 scales to extract the texture response map of the mold spot region; wherein the direction parameters are 0°, 45°, 90° and 135°, and the scale parameter σ is 1.0, 2.0 and 3.0.
[0029] Preferably, the bulb epidermis retention loss is constructed in the following manner:
[0030] Obtain the mask of the bulb epidermis region of the input image;
[0031] K-means clustering was performed on the channel vectors of the feature maps output by the teacher network and the student network in the third network stage in the epidermal region to construct a soft-assignment histogram.
[0032] The cross-entropy between the soft-assignment histograms of the teacher network and the student network is used as the bulb epidermis retention loss.
[0033] Preferably, in steps S2 and S4, the spatial attention maps are obtained by calculating the squared L2 norm along the channel direction on the specified intermediate layer feature map; step S2 uses the L2 loss between spatial attention maps as a supervision signal for parameter initialization, and step S4 uses the mean square error between spatial attention maps as the distillation loss.
[0034] Preferably, the pruning threshold is dynamically determined based on a preset ratio of the maximum absolute value of the scaling factors of all channels.
[0035] Preferably, the bulb mold patch area mask and the bulb epidermal area mask are obtained through annotation by plant pathology experts, or obtained through automatic segmentation by a lightweight semantic segmentation model and then verified by experts.
[0036] A visual recognition system for bulbous plants, deployed on the edge computing device of a 5G-connected smart horticulture vehicle, includes:
[0037] The image acquisition module is used to acquire images of bulbous plants;
[0038] The edge computing unit is equipped with a lightweight model obtained by any of the above-mentioned distillation-sensing structured channel pruning methods, which is used to perform real-time tasks such as bulb mold detection, maturity judgment and obstacle avoidance.
[0039] The 5G communication module is used for data interaction with the cloud;
[0040] The implementing agency performs obstacle avoidance or harvesting actions based on the identification results.
[0041] Compared with the prior art, the present invention has the following beneficial effects:
[0042] (1) Existing structured pruning methods such as Network Slimming have shown significant accuracy collapse when the compression rate exceeds 80%. This invention can still maintain the ability to identify tiny mold spots when the compression rate exceeds 90%; and by initializing the student network through attention transfer and combining it with knowledge distillation for pruning, the pruning process is directly guided by the teacher network's representation ability. Even in extreme cases where the compression rate exceeds 80% or even reaches 90%, the decrease in classification or detection accuracy can still be controlled within 2 percentage points.
[0043] (2) A bulb plant feature preservation loss customized for bulb plant scenarios is introduced. This loss constrains leaf color gradient, mold texture and bulb epidermal features respectively, forming a multi-level feature preservation mechanism from low level to high level. This enables the learnable scaling factor to accurately preserve channels that are sensitive to small mold spots, delicate epidermal texture and gradient leaf color. This effectively solves the problem of losing fine-grained discriminative features in general pruning methods and significantly improves the robustness of the compressed model in actual deployment scenarios.
[0044] (3) By adopting structured pruning based on channel scaling factor, the complete convolution channels are directly removed to obtain a regular and dense lightweight network structure. Significant computational acceleration can be achieved on general edge hardware without the need for special sparse library support, and the single-frame inference latency can be reduced to less than 20 milliseconds.
[0045] (4) The entire compression process is implemented end-to-end by jointly optimizing the task loss, distillation loss, sparse regularization and bulb plant feature preservation loss, avoiding suboptimal solutions caused by multi-stage independent processing, and achieving the best balance between perception accuracy, model size and inference efficiency. Attached Figure Description
[0046] Figure 1 This is a flowchart illustrating the overall implementation of the method of the present invention;
[0047] Figure 2 A schematic diagram of the training framework for distillation-sensory structured channel pruning;
[0048] Figure 3 A schematic diagram of the computational module structure for preserving the characteristics of bulbous plants.
[0049] Figure 4 This is a schematic diagram comparing the number of channels in the convolutional layers of the network before and after pruning.
[0050] Figure 5 This is an example of labeling the bulb mold area, bulb epidermis area, and leaf color gradient area in a visual task involving bulbous plants. Detailed Implementation
[0051] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.
[0052] like Figure 1 As shown, this embodiment takes the task of identifying bulb diseases and determining maturity on a 5G-connected intelligent bulb plant horticulture vehicle as an example. The teacher network is a ResNet-50 pre-trained on ImageNet and fine-tuned to convergence on the bulb plant dataset. The edge computing platform is an embedded device. The requirements are to compress the number of model parameters and computational cost to within 5% of the original model, i.e., a compression rate of over 90%, while the recognition accuracy decreases by no more than 2 percentage points, and the single-frame inference latency is less than 20 milliseconds.
[0053] Step S1: Construct the teacher network and the initial student network
[0054] On a cloud GPU cluster, ResNet-50 was trained using a dataset containing images of 10 types of bulb diseases, including mold, epidermal and leaf diseases, as well as images of healthy bulbs, epidermal images of bulbs at different maturity levels, and plant leaf images, to obtain a high-precision teacher network T.
[0055] An initial student network S0 is constructed, with a basic structure similar to ResNet-50, but the number of output channels in each bottleneck layer is uniformly reduced to 50% of the corresponding layer in the teacher network. For example, if the first bottleneck layer of the teacher network has 256 output channels, the student network is initially set to 128. At this point, the total number of parameters is approximately 25% of that of the teacher network, which does not yet meet the compression requirements. Further compression is then achieved through channel pruning.
[0056] Step S2: Attention Transfer Initialization of Student Network
[0057] Images of bulbous plants (including mold spots, epidermis, and leaves) are simultaneously input into the teacher network T and the initial student network S0. Stage 2, Stage 3, and Stage 4 output feature maps of T and S0 are selected, and attention maps along the channel dimension are calculated for each (the spatial attention map is obtained by calculating the squared L2 norm of the feature map along the channel direction). Using the teacher network's attention map as the target, the student network S0 is iteratively trained several times by minimizing the L2 loss between attention maps, causing its convolutional kernel parameters to converge quickly to an initial state that effectively focuses on key areas of mold spots, epidermal texture, and leaves. This process does not require labels and relies solely on the teacher network's hidden knowledge for transfer.
[0058] Step S3: Introduce a learnable channel scaling factor and apply sparsity regularization
[0059] like Figure 2 As shown, a learnable channel scaling factor is inserted after each convolutional layer of the student network S0. Where C is the number of output channels. During inference, the c channels of the output feature map Y are scaled to... .
[0060] During the training process, all weights Apply L1 sparse regularization term ,in Set to 1×10 -5 Sparse regularization will push the scaling factors of many unimportant channels close to zero, providing a basis for subsequent pruning.
[0061] Step S4: Construct a joint optimization objective that includes the loss of bulb plant feature retention.
[0062] The joint loss function is defined as:
[0063]
[0064] in, The cross-entropy classification loss for student networks based on disease categories or maturity levels; For attention distillation loss, the L2 distance between the attention maps of the teacher and student networks on the output feature maps of stage 2, stage 3, and stage 4 is used. The attention map is obtained by calculating the squared L2 norm of the feature vectors along the channel direction, and the weights are... Set to 1000; For sparse regularization, Set it to 1. To preserve the characteristics of bulbous plants, weighting Set it to 0.1.
[0065] like Figure 3 As shown, the specific design of the bulb plant feature retention loss L_bulb is as follows:
[0066] (1) Loss of preservation of the texture of mold spots on the bulb The binary mold region mask for each training image is obtained by using a pre-annotated bulb mold segmentation mask by plant pathologists, or by using a lightweight U-Net mold segmentation model trained on a disease dataset. Feature maps output from the teacher and student networks (stage 3) are selected, and the mold region features are first extracted using a mask. These features are then convolved using a Gabor filter bank with four directions (0°, 45°, 90°, 135°) and three scales (σ=1.0, 2.0, 3.0) to obtain 12 texture response maps. The Gram matrix of the response map within the mold region is then calculated. and ,but ².
[0067] (2) Loss of bulb epidermis retention For scenarios involving bulb maturity assessment, a mask for the bulb epidermal region can be obtained in a similar manner. K-means clustering is performed on the channel vectors of the stage4 feature maps within the epidermal region to construct a soft-assignment histogram (such as a color-texture joint histogram). The cross-entropy of the histograms of teacher and student network features in the same region is used as the basis for the assignment. Maintain sensitivity to maturity-related information such as the color and texture of the bulb's skin.
[0068] (3) Loss of leaf color gradient preservation For leaf color gradients, the Sobel operator is first used to calculate the gradient magnitude of the input image, and regions with gradient magnitudes greater than a certain threshold are selected as masks for significant leaf color gradient regions. The single-channel features are obtained by globally averaging the stage 2 output feature maps of the teacher network and student network along each channel. Calculate the gradient magnitude weighted average vector within the region, and use the L2 distance to measure the difference between the two, forming... .
[0069] Finally, the above three losses were weighted. , , The merging of these elements constitutes a loss in the retention of total bulb plant characteristics. The three types of losses guide the learning of channel scaling factors and pruning decisions. These three types of losses jointly ensure that sensitive channels of the bulb plant (mold texture, epidermal features, and leaf color gradient) are not erroneously suppressed during compression. Specifically, the bulb plant feature preservation loss constrains the leaf color gradient in stage 2, extracts the texture response map of the mold region using a Gabor filter bank in stage 3 and constrains mold texture preservation through the Frobenius norm between the Gram matrices of the teacher and student networks in stage 3, and constrains the bulb epidermal features in stage 4, forming a multi-level feature preservation mechanism from low to high levels.
[0070] Step S5: Sparse Training and Structured Pruning
[0071] Using the SGD optimizer with momentum, a joint loss is applied to the student network with scaling factor on the bulb plant dataset. Sparse training was performed with an initial learning rate of 0.01 and cosine annealing decay, for a total of 120 epochs. After training, the scaling factor for all channels was calculated. The value distribution. Set the pruning threshold to all. The absolute value is 0.1% of the maximum value; this threshold can be dynamically adjusted according to the target compression ratio. The higher the compression ratio requirement, the larger the threshold value should be. In practice, it can be selected within the range of 0.05% to 0.5%. All convolutional channels with γ below this threshold are directly removed from the network, and the input dimension of subsequent layers is adjusted simultaneously. Since the pruned object is a complete channel, the resulting model has a regular dense structure with no sparse matrix computation overhead.
[0072] Taking this embodiment as an example, the changes in the number of channels before and after pruning at each stage are as follows: Conv1 decreases from 64 to 16, Layer1 decreases from 256 to 32, Layer2 decreases from 512 to 64, Layer3 decreases from 1024 to 128, and Layer4 decreases from 2048 to 256.
[0073] like Figure 4 As shown, taking this embodiment as an example, after pruning, the number of parameters in the initial student network is reduced to about 5% of that in the teacher network, and the computational load is also reduced to about 5%.
[0074] Step S6: Fine-tune the lightweight model and deploy it.
[0075] The pruned, compact model was retrained for 30 epochs on the same bulbous plant dataset with a small learning rate (0.001) to recover the slight accuracy loss caused by channel deletion. The fine-tuned model was then converted to an edge-device-optimized inference engine via TensorRT and run at FP16 accuracy.
[0076] like Figure 5 As shown in the field tests, the model's Top-1 accuracy is only about 2% lower than the teacher network in tasks such as bulb mold spot recognition, bulb maturity judgment, and leaf disease detection; the single-frame inference latency is about 20ms, meeting the requirements for real-time perception. Meanwhile, downstream task indicators such as maturity judgment related to bulb epidermal texture and leaf color gradation remain good, fully verifying the ability of the proposed distillation-sensing structured channel pruning method to preserve fine-grained features of bulbous plants under extremely high compression ratios.
[0077] In summary, addressing the issues of excessively large models and high inference latency faced by 5G-connected intelligent bulb plant horticulture vehicles on edge computing devices, this invention proposes a compression method combining joint knowledge distillation, structured pruning, and bulb plant feature preservation loss. First, the student network is initialized using an attention transfer mechanism. Then, a learnable channel scaling factor is introduced, and sparse regularization is applied to construct a joint loss function for sparse training, which includes task loss, distillation loss, sparse regularization loss, and bulb plant feature preservation loss. This feature preservation loss constrains leaf color gradients, mold texture using Gabor filter banks and Gram matrices, and bulb epidermal features through a multi-level feature preservation mechanism, maintaining feature consistency before and after pruning from low to high levels. Finally, low-contribution channels are pruned based on the scaling factor, and fine-tuning is performed. The beneficial effects of this invention are that it maintains a recognition accuracy decrease of no more than 2 percentage points and an inference latency of less than 20ms while achieving a compression rate of over 90%, making it directly applicable to large-scale cross-modal models in intelligent horticulture on edge devices.
[0078] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.
Claims
1. A distillation-based structured channel pruning method for visual recognition of bulbous plants, characterized in that, Includes the following steps: Step S1: Obtain the teacher network trained to convergence on the target task dataset, and construct the initial student network with a reduced number of channels; Step S2: Using the attention transfer mechanism, the spatial attention map corresponding to the intermediate layer feature map specified by the teacher network is used as the supervision target to initialize the parameters of the initial student network; Step S3: Introduce a learnable channel scaling factor for each convolutional layer of the student network and apply a sparse regularization term during training; Step S4: Construct a joint loss function to perform sparse training on the student network containing the channel scaling factor; the joint loss function includes at least the task loss, the distillation loss for aligning the spatial attention maps corresponding to the feature maps of the teacher network and the student network in the specified intermediate layer, the sparse regularization loss applied to the channel scaling factor, and the bulb plant feature preservation loss for maintaining the feature consistency of the teacher network and the student network in the plant sensitive region; Step S5: After sparse training is completed, based on the size of the channel scaling factor, the entire convolutional channel with a scaling factor lower than the pruning threshold is pruned to obtain the structured pruning model. Step S6: Fine-tune the structured pruning model using the target task dataset to obtain a lightweight model for visual recognition of bulbous plants.
2. The distillation-sensory structured channel pruning method for visual recognition of bulbous plants according to claim 1, characterized in that, The bulb plant feature preservation loss includes: leaf color gradient preservation loss applied to the first network stage output of the teacher network and student network, bulb mold texture preservation loss applied to the second network stage output, and bulb epidermis preservation loss applied to the third network stage, forming a multi-level feature preservation mechanism from low to high level.
3. The distillation-sensory structured channel pruning method for visual recognition of bulbous plants according to claim 2, characterized in that, The leaf color gradient preservation loss is constructed in the following way: Calculate the gradient magnitude of the input image, and take the region with a gradient magnitude greater than a preset threshold as the mask for the region with significant leaf color gradient; The feature maps output by the teacher network and the student network in the first network stage are globally averaged along the channel direction to obtain single-channel feature maps. Calculate the weighted average vector of the gradient magnitude of the single-channel feature map within the mask region, and use L2 distance to constrain the weighted average vectors of the teacher network and the student network to tend to be consistent.
4. The distillation-sensory structured channel pruning method for visual recognition of bulbous plants according to claim 2, characterized in that, The bulb mold texture retention loss is constructed in the following way: Obtain the mask of the bulb mold patch region in the input image; For the feature maps output from the second network stage of the teacher network and the student network, the features of the moldy area are extracted separately using the bulb moldy area mask. The extracted mold patch region features were convolved using a Gabor filter bank to obtain multiple sets of texture response maps; Calculate the Gram matrix of the texture response map within the mold patch region; The difference between the Gram matrices of the teacher network and the student network is measured using the Frobenius norm, and is used as the bulb mold texture preservation loss.
5. The distillation-sensory structured channel pruning method for visual recognition of bulbous plants according to claim 4, characterized in that, The texture preservation loss of the bulb mold spot was obtained by extracting the texture response map of the mold spot region using a Gabor filter bank with 4 directions and 3 scales; where the direction parameters are 0°, 45°, 90° and 135°, and the scale parameter σ is 1.0, 2.0 and 3.
0.
6. The distillation-sensory structured channel pruning method for visual recognition of bulbous plants according to claim 2, characterized in that, The bulb epidermal retention loss is constructed in the following way: Obtain the mask of the bulb epidermis region of the input image; K-means clustering was performed on the channel vectors of the feature maps output by the teacher network and the student network in the third network stage in the epidermal region to construct a soft-assignment histogram. The cross-entropy between the soft-assignment histograms of the teacher network and the student network is used as the bulb epidermis retention loss.
7. The distillation-sensory structured channel pruning method for visual recognition of bulbous plants according to claim 1, characterized in that, In steps S2 and S4, the spatial attention maps are obtained by calculating the squared L2 norm along the channel direction on the specified intermediate layer feature map; step S2 uses the L2 loss between spatial attention maps as a supervision signal for parameter initialization, and step S4 uses the mean square error between spatial attention maps as the distillation loss.
8. The distillation-sensory structured channel pruning method for visual recognition of bulbous plants according to claim 1, characterized in that, The pruning threshold is dynamically determined based on a preset ratio of the maximum absolute value of the scaling factors of all channels.
9. The distillation-sensory structured channel pruning method for visual recognition of bulbous plants according to claim 4 or 6, characterized in that, The bulb mold patch area mask and the bulb epidermal area mask were obtained through annotation by plant pathology experts, or obtained through automatic segmentation by a lightweight semantic segmentation model and then verified by experts.
10. A visual recognition system for bulbous plants, deployed on the edge computing device of a 5G-connected intelligent horticulture vehicle, characterized in that, include: The image acquisition module is used to acquire images of bulbous plants; The edge computing unit is equipped with a lightweight model obtained by the distillation-sensing structured channel pruning method described in any one of claims 1 to 8, and is used to perform real-time tasks such as bulb mold detection, maturity judgment and obstacle avoidance. The 5G communication module is used for data interaction with the cloud; The implementing agency performs obstacle avoidance or harvesting actions based on the identification results.