CN-121999514-A - Wheat aphid hazard assessment method based on improved YOLO instance segmentation model

Abstract

The invention discloses a wheat aphid hazard assessment method based on an improved YOLO instance segmentation model, which comprises the steps of S1, preprocessing an instance perception image, S2, constructing an instance segmentation model, carrying out structural improvement, introducing a plurality of EV blocks as convolution modules and embedding the EV blocks into an FMCA module, carrying out model training, and S3, outputting a result, and calculating a plurality of quantitative indexes for assessment. The method improves a YOLO instance segmentation network structure, better captures global features to solve the problem of compatibility of micro targets and large-scale targets, improves a slicing process, combines dynamic cutting and regular grid cutting, introduces a geometric topology restoration and negative sample probability retention mechanism, improves the quality of training samples of the micro targets in high-resolution images, constructs a multidimensional quantitative evaluation index system, and realizes unified quantitative characterization of the multiple hazard degrees of wheat aphids by means of nonlinear function fusion density, coverage and aggregation.

Inventors

• Deng Haotian
• FAN SHUMING
• Pei Zijun
• SUN QI
• HUANG MEI
• WANG XIAO
• ZHOU QIN
• ZHONG YINGXIN
• CAI JIAN
• LIU XIN
• LI QING
• JIANG DONG
• CHEN JIAWEI
• Pan Qiuxiao
• ZHANG ZHEN
• Qin Bingxi
• ZHOU HONGHAO

Assignees

• 南京农业大学
• 南京慧瞳作物表型组学研究院有限公司
• 河南大学
• 山东和点农业科技有限公司

Dates

Publication Date

20260508

Application Date

20251224

Priority Date

20251219

Claims (10)

1. 1. A wheat aphid hazard assessment method based on an improved YOLO instance segmentation model is used for processing an instance perceived image of wheat and is characterized by comprising the following steps: S1, preprocessing an instance perceived image, wherein the preprocessing comprises dual-path slicing generation, geometric topology restoration and negative sample balance control; S2, constructing an EVC-YOLO-Seg example segmentation model, and carrying out structural improvement and training aiming at a YOLO architecture; The method comprises the steps of carrying out structural improvement, introducing a plurality of EV blocks into a backbone network of a YOLO architecture to serve as convolution modules, embedding an FMCA module between a Neck network output end of the YOLO architecture and a Head network, adopting Shape-IoU as a boundary frame regression loss function during training, and carrying out model training by using a marked data set; S3, inputting the preprocessed instance perceived image into a trained EVC-YOLO-Seg instance segmentation model to obtain an instance segmentation result, and calculating a plurality of quantitative indexes for evaluation based on the instance segmentation result.
2. 2. The method for wheat aphid hazard assessment based on the improved YOLO instance segmentation model as set forth in claim 1, wherein the dual-path cut generation in step S1 comprises true-value based focus cutting and regular grid sliding window based cutting for generating a plurality of cuts.
3. 3. The wheat aphid hazard assessment method based on the improved YOLO instance segmentation model according to claim 2, wherein in the step S1, geometric topology restoration comprises the steps of performing geometric intersection calculation on polygonal labels corresponding to any cut with truncated edges, and executing topology restoration when self-intersecting or degraded edges occur, wherein before geometric topology restoration, invalid cut pieces with the area smaller than a set threshold or vertex number smaller than E are eliminated from a plurality of cut pieces obtained by generating double-path cut pieces, and E is a positive integer.
4. 4. The wheat aphid hazard assessment method based on the improved YOLO instance segmentation model of claim 1, wherein the negative sample balance control in step S1 comprises setting a null slice retention probability, and rejecting a plurality of pure background slices by a bernoulli sampling mechanism.
5. 5. The wheat aphid hazard assessment method based on the improved YOLO instance segmentation model of claim 1, wherein in step S2, any EV Block includes an HSM-SSD unit that performs feature extraction on long-range dependence of aphid features using an SMM state space model.
6. 6. The wheat aphid hazard assessment method based on the improved YOLO instance segmentation model of claim 1, wherein in step S2, the structure of any EV Block is a FFN feed-forward network, DWConv depth separable convolution, HSM-SSD unit, DVConv depth deformable convolution connected in sequence.
7. 7. The wheat aphid hazard assessment method based on the improved YOLO instance segmentation model as claimed in claim 1, wherein the FMCA module in step S2 fuses CAM channel attention, SAM spatial attention and MSCA multi-scale convolution attention, which accumulates context information using multi-scale strip rolls.
8. 8. The wheat aphid hazard assessment method based on the improved YOLO instance segmentation model as claimed in claim 1, wherein a Shape-IoU of step S2 introduces a Shape weight coefficient sensitive to direction, and the penalty weight is adjusted according to the difference of the horizontal and vertical dimensions of the prediction frame and the real frame.
9. 9. The wheat aphid hazard assessment method based on the improved YOLO instance segmentation model as claimed in claim 1, wherein the HSM-SSD unit comprises a plurality of Linear full connection layers, a plurality of MatMul matrix multiplication units, a depth separable convolution unit, a Discretization discretization unit and a plurality of Lin Linear layers.
10. 10. The method for wheat aphid hazard assessment based on the improved YOLO instance segmentation model of claim 1, wherein in step S3, the plurality of quantitative indicators comprises at least a number, a density, a coverage, a concentration, a cluster feature, and a comprehensive hazard indicator ADI.

Description

Wheat aphid hazard assessment method based on improved YOLO instance segmentation model Technical Field The invention relates to the technical field of intelligent monitoring and computer vision image processing of agricultural diseases and insect pests, in particular to a wheat aphid hazard assessment method based on an improved YOLO instance segmentation model. Background At present, the wheat aphid monitoring and hazard assessment method is mainly divided into two major categories, namely traditional manual investigation and automatic monitoring based on computer vision. Traditional methods rely primarily on manual field surveys and pest trapping stations. The manual investigation has high accuracy, but has high labor intensity, low efficiency and subjective error, and the trapping station can monitor in real time, but lacks mobility, and is difficult to cover a large area or cope with a dynamic monitoring scene. With the development of deep learning, a target detection technology based on a convolutional neural network has been applied to crop pest monitoring. However, the existing evaluation method mainly focuses on the index of counting the number of the pests in a single dimension, namely, the scale of the pests is deduced through the number of the detection frames, a multi-dimension evaluation pest system is not constructed, and the image data cannot be fully utilized for carrying out multi-dimension comprehensive quantitative pest evaluation. Despite the progress made in the prior art, the following main problems and disadvantages still exist in the aspects of accurate segmentation and multidimensional quantitative evaluation of wheat aphids: (1) The tiny dense target features are lost and interfere seriously with the background. The wheat aphid individuals are tiny and densely distributed, and the colors are highly similar to the leaf background, so that the detection omission and false detection are very easy to occur under the complex field background (such as dew, honeydew and impurity interference). Analysis of the cause traditional neural network models often take the whole image directly as input or do simple downsampling when processing high resolution images. Fine-grained features of tiny objects are easily lost in deep networks due to multiple downsampling and feature map resolution limitations. In addition, existing models lack feature enhancement mechanisms for micro-texture and low contrast targets, making it difficult to distinguish aphids from similar background noise. (2) Extreme unbalance of the size target dimensions results in limited segmentation accuracy. In wheat aphid hazard assessment, it is necessary to divide both extremely tiny aphids and larger leaves simultaneously to calculate density. The existing model is difficult to consider the targets of the two extreme scales, and the situation that the blade is not completely segmented or aphid is not completely segmented often occurs. The traditional convolution layer has a fixed receptive field, and is difficult to pay attention to the detail characteristics of a small target and the global structure information of a large target at the same time. Furthermore, the class imbalance of "huge aphid count" versus "rare leaf count" in the dataset can lead to gradient bias in the model optimization process, making the model prone to overfitting the high frequency class (aphid) while ignoring the low frequency class (leaf). (3) The high-precision model has high computational complexity and is difficult to deploy at a mobile terminal. The existing high-performance example segmentation model (such as Mask R-CNN, mask2 Former) is huge in parameter quantity (such as the size of Mask2Former model is 780.5 MB), high in calculation cost (FLOPs high) and low in reasoning speed (only about 6 FPS), and cannot meet the requirement of field mobile equipment real-time monitoring. The reason analysis is that the models adopt complex two-stage structures or heavy-duty architectures based on transformers, and are not designed in a light-weight way aiming at the computational power limitation of a mobile terminal. However, the existing lightweight model (such as Fast-SCNN) is Fast, but the precision of dividing the boundary of a tiny target under a complex background is often seriously lost. (4) The hazard evaluation index is single, and quantitative characterization of ecological dimension is lacking. Existing automated monitoring techniques rely primarily on quantitative indicators to evaluate pests. Such a single index cannot reflect the coverage area of the pest (affecting the photosynthesis level), the spatial aggregation distribution pattern (reflecting the population diffusion trend), and the local burst point characteristics. The reason analysis is that most of the prior art stops at a target detection task, and cannot further utilize pixel level information provided by an example segmentation algorithm to construct a multi-dimensional compre