Moso bamboo age identification method and system based on improved YOLO11 model

Abstract

The invention discloses a moso bamboo age identification method and system based on an improved YOLO11 model, and belongs to the crossing field of artificial intelligence and forestry resource monitoring. Aiming at the problems in the prior art that bamboo age identification depends on artificial experience, image processing is easily interfered by illumination, and a general model is insufficient in cross-scale texture feature capture and the like, the invention provides the following innovations: 1) an anti-reflection-texture decoupling fusion framework is designed, reflection high-frequency noise is eliminated through a GhostConv module, longitudinal textures are enhanced in combination with a CBAM channel-space attention mechanism, and the cross-scale texture features are not sufficiently captured; a C2PSA multi-scale pyramid is used for fusing the features under different illumination conditions; 2) constructing a multi-granularity age sensitive detection head, and extracting bamboo joint spacing features in a cross-scale manner by adopting a cavity convolution combination (1 * 1 / 3 * 3 / 5 * 5) of a Basic RFB module; and 3) introducing lightweight collaborative design, compressing the volume of the model to 12.5 M (42% less than that of YOLOv5s) by using depth separable convolution (kernel size = 2) of a C3k2 module and a GhostConv channel compression technology, and keeping edge sharpness at the same time. Through verification of 2086 annotated images in three places, the age identification accuracy of the method reaches 89.5% and is improved by 3.8% compared with a base line, 120ms real-time detection of a mobile terminal is supported, and the method can be efficiently used for bamboo forest resource investigation and dynamic monitoring.

Description

Technical Field

[0001] The present invention belongs to the intersection of artificial intelligence and forestry resource monitoring, and specifically relates to a computer vision method and system based on an improved YOLO11 deep learning model, which enhances the semantic information of the surface texture, color and morphological characteristics of the bamboo stems in RGB images to achieve high-precision and low-parameter identification of the age of bamboo. Background Art

[0002] 1. Deficiencies in existing technology

[0003] Traditional method: relies on manual observation of bamboo stem skin color, internode length and other characteristics, which is inefficient and highly subjective (error rate >30%).

[0004] Image processing technology: The color analysis method based on threshold segmentation (such as HSV spatial age classification) is easily affected by the lighting environment, and the accuracy rate of bamboo age recognition is less than 50%.

[0005] General object detection model: When the YOLO series model is directly applied to bamboo age recognition, the mAP is lower than 70% due to the difficulty in capturing the subtle differences in bamboo stem texture and complex background interference (such as dead leaves and moss in the forest).

[0006] 2. Technical bottleneck

[0007] Bamboo age characteristics have cross-scale properties: for young bamboo (grade I bamboo), attention should be paid to the micro-texture of the epidermal hairs, while for mature bamboo (grade II to grade IV bamboo), the macro-characteristics of the bamboo node morphology should be analyzed.

[0008] Data acquisition limitations: The ambient lighting environment under the forest is complex, and the reflective properties of the bamboo pole surface cause large fluctuations in conventional RGB image information. Summary of the Invention

[0009] Core innovations

[0010] 1. Anti-reflection-texture decoupling fusion architecture

[0011] Technical features:

[0012] GhostConv+CBAM+C2PSA three-stage feature processing module

[0013] GhostConv (layers 1 / 4 / 6 / 9): removes high-frequency noise in reflective areas through linear feature compression

[0014] CBAM (layer 3 / 8): Strengthening bamboo joint longitudinal texture response based on channel-spatial attention mechanism

[0015] C2PSA (layer 12): multi-scale pyramid feature aggregation, fusing texture features under different lighting conditions

[0016] 2. Multi-granularity age-sensitive detection head

[0017] BasicRFB+dynamic receptive field combination structure (layers 19 / 23 / 27)

[0018] BasicRFB module (layers 19 / 23 / 27): captures the longitudinal internode features of bamboo culms through a combination of dilated convolutions (1×1, 3×3, 5×5)

[0019] 3. Lightweight-high-precision collaborative design

[0020] Compression-activation mechanism of C3k2+GhostConv combination

[0021] C3k2 (layers 2 / 5 / 7 / 10): uses depthwise separable convolution with kernel_size=2, reducing the number of parameters by 50% while maintaining the sharpness of the edge of the node

[0022] GhostConv: Eliminates redundant feature channels in key Backbone layers (layers 1 / 4 / 6 / 9), reducing the model size to 12.5MB (42% smaller than the original YOLOv5s)

[0023] Technical Effects

[0024] In the test of a verification set of pictures taken in Yong'an, Zhengzhou and Nanjing in Fujian (including 2,086 annotated images), the age recognition accuracy reached 89.5%, an increase of 3.8% over the baseline YOLO11.

[0025] The lightweight version of the model can achieve 120ms real-time detection on Kirin 985 mobile devices, meeting the needs of field deployment.

[0026] 1. Data Collection and Preprocessing

[0027] The bamboo stem RGB (400-700 nm) images were collected using a smartphone camera with a resolution of ≥12 million pixels.

[0028] Marking specification: Take the three nodes at 1.3m from the base of the bamboo pole as the ROI and mark them according to the actual felling age (grades I-IV).

[0029] 2. Model training details

[0030]

[0031] 3. Deployment plan

[0032] BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1:System flow chart (data collection → improved YOLO11 reasoning → age output)

[0034] Figure 2 : Improved YOLO11 model architecture diagram

[0035] Figure 3 :CBAM module

[0036] Figure 4 :BaiscBRF module

[0037] Figure 5 : Characteristics of bamboo poles of different ages.

Claims

1. A method for identifying bamboo age based on an improved YOLO11 model, characterized in that: The following steps are involved: Anti-reflection and texture decoupling fusion processing: Through three-stage feature processing of the GhostConv module, the CBAM module, and the C2PSA module, the bamboo pole reflective noise is eliminated and the texture features are enhanced; The GhostConv module is deployed in network layers 1 / 4 / 6 / 9 to eliminate high-frequency noise in reflective areas through linear feature compression; The CBAM module is deployed at network layers 3 / 8, enhancing the bamboo joint longitudinal texture response through a channel-spatial attention mechanism. The C2PSA module is deployed in network layer 12 and fuses texture features under different illuminations through multi-scale pyramid feature aggregation; Multi-granularity age-sensitive detection: The BasicRFB module is combined with a dynamic receptive field structure to capture the longitudinal internode characteristics of bamboo stems. The BasicRFB module is deployed in network layers 19 / 23 / 27 and uses a combination of dilated convolutions (1x1, 3x3, and 5x5) to achieve cross-scale feature extraction. Lightweight-high-precision collaborative optimization: The model parameters are compressed by combining the C3k2 module and the GhostConv module. The C3k2 module uses depthwise separable convolution with kernel_size = 2 and is deployed at network layers 2 / 5 / 7 / 10, maintaining the sharpness of node edges while reducing the number of parameters.

2. The method according to claim 1, characterized in that The GhostConv module eliminates feature channel redundancy and compresses the model volume to 12.5M, which is 42% smaller than the original YOLOv5s model.

3. The method according to claim 1, characterized in that The data preprocessing includes: Three sections of the bamboo stem at 1.3 m from the base were collected as ROIs and the actual felling age (grades I-IV) was marked; The input image resolution is ≥12 million pixels, and data augmentation is performed in the HSV space, with a hue adjustment probability of 0.015, a saturation adjustment probability of 0.7, and a lightness adjustment probability of 0.

4.

4. The method according to claim 1, wherein The model training parameters are configured as follows: The number of iterations is 300, the batch size is 16, the learning rate is 0.01, the momentum coefficient is 0.937, and the weight decay coefficient is 0.0005; The weight coefficients of target loss, classification loss, and bounding box loss are 1.5, 7.5, and 7.5, respectively.

5. The method according to claim 1, wherein When deployed, the model has 2,720,184 parameters, 6.5 GFLOPs, and a detection speed of 120ms / frame on mobile devices.

6. The method according to claim 1, characterized in that The C2PSA module aggregates multi-scale pyramid features and fuses texture semantic information under different lighting conditions to improve the robustness of cross-age features.

7. A bamboo age recognition system based on an improved YOLO11 model, characterized in that: include: Anti-reflection-texture decoupling fusion module: It consists of a GhostConv module, a CBAM module, and a C2PSA module in series, and is used to eliminate reflective noise and enhance texture features; Multi-granularity detection head module: This module includes the BasicRFB module and the dynamic receptive field structure, which is used to extract cross-scale features of the longitudinal internode spacing of bamboo stems; Lightweight compression module: It consists of a C3k2 module and a GhostConv module in parallel, which is used to reduce the number of model parameters and maintain detection accuracy; Deployment interface module: supports real-time inference on mobile devices with Kirin 985 and above, and outputs bamboo age classification results (degrees I-IV).

8. The system according to claim 7, characterized in that The C3k2 module adopts depthwise separable convolution with kernel_size=2, which reduces the number of parameters while retaining the sharpness of the edge details.

Images