CN116524255A - Wheat scab spore identification method based on Yolov5-ECA-ASFF - Google Patents
Wheat scab spore identification method based on Yolov5-ECA-ASFF Download PDFInfo
- Publication number
- CN116524255A CN116524255A CN202310444199.6A CN202310444199A CN116524255A CN 116524255 A CN116524255 A CN 116524255A CN 202310444199 A CN202310444199 A CN 202310444199A CN 116524255 A CN116524255 A CN 116524255A
- Authority
- CN
- China
- Prior art keywords
- spore
- wheat scab
- layer
- asff
- spores
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 206010039509 Scab Diseases 0.000 title claims abstract description 106
- 241000209140 Triticum Species 0.000 title claims abstract description 92
- 235000021307 Triticum Nutrition 0.000 title claims abstract description 92
- 238000000034 method Methods 0.000 title claims abstract description 40
- 238000001514 detection method Methods 0.000 claims abstract description 72
- 230000004927 fusion Effects 0.000 claims abstract description 33
- 238000012549 training Methods 0.000 claims abstract description 32
- 238000000605 extraction Methods 0.000 claims abstract description 11
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims abstract description 6
- 238000013507 mapping Methods 0.000 claims description 26
- 230000006870 function Effects 0.000 claims description 25
- 238000005070 sampling Methods 0.000 claims description 19
- 238000004364 calculation method Methods 0.000 claims description 17
- 241000233866 Fungi Species 0.000 claims description 14
- 230000008569 process Effects 0.000 claims description 7
- 238000011156 evaluation Methods 0.000 claims description 6
- 238000002372 labelling Methods 0.000 claims description 6
- 238000011176 pooling Methods 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 6
- 230000002708 enhancing effect Effects 0.000 claims description 5
- 239000013598 vector Substances 0.000 claims description 5
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 238000009825 accumulation Methods 0.000 claims description 3
- 238000005520 cutting process Methods 0.000 claims description 3
- 238000009826 distribution Methods 0.000 claims description 3
- 230000002068 genetic effect Effects 0.000 claims description 3
- 238000012886 linear function Methods 0.000 claims description 3
- 230000003044 adaptive effect Effects 0.000 claims description 2
- 230000007246 mechanism Effects 0.000 abstract description 15
- 238000010586 diagram Methods 0.000 abstract description 13
- 230000007547 defect Effects 0.000 abstract description 2
- 238000002474 experimental method Methods 0.000 description 12
- 230000000694 effects Effects 0.000 description 8
- 238000013135 deep learning Methods 0.000 description 6
- 201000010099 disease Diseases 0.000 description 5
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 5
- 239000011521 glass Substances 0.000 description 5
- 239000001963 growth medium Substances 0.000 description 5
- 238000010801 machine learning Methods 0.000 description 5
- 230000002538 fungal effect Effects 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 239000007787 solid Substances 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 230000000052 comparative effect Effects 0.000 description 3
- 238000005286 illumination Methods 0.000 description 3
- 230000003993 interaction Effects 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- UHPMCKVQTMMPCG-UHFFFAOYSA-N 5,8-dihydroxy-2-methoxy-6-methyl-7-(2-oxopropyl)naphthalene-1,4-dione Chemical compound CC1=C(CC(C)=O)C(O)=C2C(=O)C(OC)=CC(=O)C2=C1O UHPMCKVQTMMPCG-UHFFFAOYSA-N 0.000 description 2
- 241000223218 Fusarium Species 0.000 description 2
- 241000223195 Fusarium graminearum Species 0.000 description 2
- 230000002457 bidirectional effect Effects 0.000 description 2
- 238000013145 classification model Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000003066 decision tree Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000001965 increasing effect Effects 0.000 description 2
- 238000001000 micrograph Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 239000000725 suspension Substances 0.000 description 2
- HPTJABJPZMULFH-UHFFFAOYSA-N 12-[(Cyclohexylcarbamoyl)amino]dodecanoic acid Chemical compound OC(=O)CCCCCCCCCCCNC(=O)NC1CCCCC1 HPTJABJPZMULFH-UHFFFAOYSA-N 0.000 description 1
- 241000123650 Botrytis cinerea Species 0.000 description 1
- 241000196324 Embryophyta Species 0.000 description 1
- 241000221785 Erysiphales Species 0.000 description 1
- 241000233732 Fusarium verticillioides Species 0.000 description 1
- 241001248610 Ophiocordyceps sinensis Species 0.000 description 1
- 240000007594 Oryza sativa Species 0.000 description 1
- 235000007164 Oryza sativa Nutrition 0.000 description 1
- 244000061456 Solanum tuberosum Species 0.000 description 1
- 235000002595 Solanum tuberosum Nutrition 0.000 description 1
- 241001251949 Xanthium sibiricum Species 0.000 description 1
- 241000607479 Yersinia pestis Species 0.000 description 1
- 238000002679 ablation Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 239000000084 colloidal system Substances 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 239000000428 dust Substances 0.000 description 1
- 235000013305 food Nutrition 0.000 description 1
- 239000007788 liquid Substances 0.000 description 1
- 230000000873 masking effect Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 244000005700 microbiome Species 0.000 description 1
- 238000002156 mixing Methods 0.000 description 1
- 230000000877 morphologic effect Effects 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 239000002245 particle Substances 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 235000009566 rice Nutrition 0.000 description 1
- 238000007790 scraping Methods 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
- G06V10/765—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects using rules for classification or partitioning the feature space
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/52—Scale-space analysis, e.g. wavelet analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
- G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Databases & Information Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Medical Informatics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a wheat scab spore identification method based on Yolov5-ECA-ASFF, which solves the defect that detection and identification are difficult to carry out on a small target of which the sample is scab spores compared with the prior art. The invention comprises the following steps: establishing a spore image data set; constructing a wheat scab spore recognition model; training a wheat scab spore recognition model; obtaining wheat scab spores to be identified; and obtaining the identification result of wheat scab spores to be identified. The invention adds ECA space attention mechanism at the end of CSPNet residual block of YOLOv5s backbone network to strengthen the channel characteristics of the input end characteristic diagram; an ASFF module with a self-adaptive feature fusion mechanism is introduced at the end of a Neck feature extraction network, so that the wheat scab spores are effectively and rapidly detected and identified.
Description
Technical Field
The invention relates to the technical field of wheat scab spore recognition, in particular to a wheat scab spore recognition method based on Yolov 5-ECA-ASFF.
Background
The spore image identification is the application of an image identification algorithm in the field of agricultural pest and disease fungi spore detection, and is mainly used for accurately positioning target disease spores on images. With the rapid development of image processing technology and artificial intelligence algorithms, spore image identification has become a new research hotspot and has achieved a certain result. Spore image recognition techniques can be classified into machine learning and deep learning methods according to different algorithm principles. In disease control, a large number of samples are required to train the classifier to achieve a classification judgment of spore types.
In machine learning spore identification research, yang et al put forward that after the separation of the adhesion spores by a distance conversion Gaussian filter algorithm method through cooperative judgment of texture and shape characteristics on a microscopic image dataset of rice spores and a decision tree fusion matrix method, decision tree model classification is carried out by four shape characteristics (area, circumference, ovality and complexity) and three texture characteristics (entropy, homogeneity and contrast), and the detection accuracy is as high as 94%. However, the method has less extraction characteristics and single target, so that the method has poor practicability. Wang et al set up SVM (Support Vector achine) classification models based on 600 datasets of botrytis cinerea spores, guba powdery mildew spores and Xanthium sibiricum spores by using image preprocessing methods such as mean filtering, gaussian filtering, maximum inter-class variance binarization, morphological operation and masking operation to extract 90 features of spores, and the results show that the accuracy of the SVM model reaches 91.68% as a total average value, but the method does not perform dimension reduction optimization on a large number of features, so that the detection accuracy is not very high and the detection rate is limited. Wang et al used 13 diffraction fingerprint feature values selected for SVM classification model based on potato disease diffraction fingerprint image dataset, test results showed that the average accuracy of identification of three fungal spores in greenhouse crops by support vector machine model was 92.72%, the accuracy was improved after feature optimization, but the accuracy was not verified when other microorganisms were present in the air.
Although the traditional machine learning has a certain achievement in spore detection, the method is only suitable for the situations of single target, obvious characteristics and simple background, and for the recognition tasks of multiple targets and complex and changeable background, the method is difficult to extract the surface characteristics by utilizing the machine learning so as to achieve good detection effect performance, so that deep learning is required to extract a large number of characteristics of the complex targets through a convolution structure to complete the recognition task.
With the improvement of the hardware performance of the computer, deep learning has been rapidly developed with the advantages of low cost and high efficiency, and various neural networks have been applied to small target detection of microscopic images. For example, jukayer et al used a pre-trained Yolov5 algorithm on microscopic images of mold on the surface of food, and the accuracy of mold detection reached 98.10%, but it was not tested in a complex background and the ability of the network to extract features for small targets was not verified. Wang et al propose a lightweight Yolov4 network based on a fusion attention mechanism ECA, local cross-channel interaction is efficiently realized by one-dimensional convolution, dependency relations among channels are extracted, recognition of small targets is improved, and detection performance difference of the network under different scenes is large. The Qiu et al propose that Yolov5 network combines with adaptive feature fusion ASFF to enhance the detection effect of small targets in view of the problems of dense targets, multiple scales and small targets in road traffic detection, but the detection effect is still wrong when the targets are blocked. Dadboud et al realize better detection performance by adopting a Yolov5 mosaic data enhancement technology aiming at the challenges of small detection targets and feature uncertainty of an unmanned aerial vehicle, but have poor robustness and poor detection effect in complex scenes.
Compared with machine learning, the deep learning has strong learning capacity, does not need a large amount of characteristic engineering, and has strong portability. However, in small target detection, the method has the characteristics of few available characteristics, high positioning accuracy requirement and the like, and the common network structure does not optimize the small target detection.
However, most of the traditional research methods further extract multidimensional features and screen features for optimization by pre-dividing target areas, and then utilize the screened features to combine with a plurality of classifiers such as SVM and the like to realize target detection; the deep learning is performed through a complex convolutional neural network structure, the optimal characteristics of different targets are adaptively extracted, redundant characteristic extraction and region pre-segmentation are not needed, detection and classification tasks can be simultaneously executed, microscopic images with more complex background environments can be detected, and higher detection precision is realized. However, deep learning has the characteristic of higher requirement on a data set, a large amount of sample size is needed as a data support to prevent the phenomena of fitting and low precision, and the complexity of a structure and the convolution calculation amount lead to higher computer power required for network training. The existing work is often only to carry out single structural modification on the network feature extraction capability or the multidimensional semantic information extraction capability, and the network detection capability under multiple scenes is poor in performance although the detection effect is improved.
Therefore, how to realize effective and rapid detection and identification of wheat scab spores has become a technical problem to be solved urgently.
Disclosure of Invention
The invention aims to solve the defect that detection and identification are difficult to be carried out on a small target of which a sample is scab spores in the prior art, and provides a wheat scab spore identification method based on Yolov5-ECA-ASFF to solve the problems.
In order to achieve the above object, the technical scheme of the present invention is as follows:
a wheat scab spore identification method based on Yolov5-ECA-ASFF comprises the following steps:
establishment of spore image dataset: collecting and collecting scab spore images, preprocessing the images and enhancing the data to construct a spore image data set;
constructing a wheat scab spore recognition model: based on a Yolov5 network, an ECA module and an ASFF module are fused to construct a wheat scab spore recognition model;
training of wheat scab spore recognition model: inputting the spore image dataset into a wheat scab spore recognition model for training;
obtaining wheat scab spores to be identified: acquiring wheat scab spores to be identified, and preprocessing;
obtaining the identification result of wheat scab spores to be identified: and inputting the pretreated wheat scab spores to be identified into a trained wheat scab spore identification model to obtain a wheat scab spore identification result.
The establishment of the spore image dataset comprises the following steps:
selecting 10000 spore image data containing 4 kinds of wheat scab and 4 kinds of mixed fungus spore image data of 4 kinds of scab fungus spores from the acquired images as a target identification data set;
expanding 10000 data sets to 20000 data sets through data enhancement;
and manually labeling the amplified picture data by using Labelme software, and converting the obtained data set with json file label type into the label file type required by Yolov5 after labeling to construct a spore image data set.
The construction of the wheat scab spore recognition model comprises the following steps:
constructing a wheat scab spore recognition model based on Yolov5, and setting the wheat scab spore recognition model into four layers:
the first layer is Input, the Input is used as an Input end for adaptively scaling an image, a K-means genetic algorithm is integrated, an optimal anchor frame value of a data set is adaptively calculated, and the small target detection capability is enhanced;
the second layer is a Backbone structure of the Backbone network, and the Backbone structure comprises a Focus, a CSP module and an SPP module; the Focus integrates the width and height information of the input photo in the channel, the CSP structure is divided into two parts by carrying out feature mapping on the base layer, a cross-stage hierarchical structure is adopted to fuse the two parts, and an SPP module at the tail end is used for three structures of a convolution layer, a pooling layer and a selection filter;
the third layer is a Neck structure which comprises an up-down sampling structure FPN layer of FPN+PAN, and the tail end is combined with a PAN module to carry out up-sampling;
the fourth layer is Output, the Output end is used for evaluating whether the detection and positioning of the algorithm on the target are accurate or not, the accuracy is calculated by using a GIOU Loss function, the GIOU Loss function eliminates the residual prediction result of dividing the GIOU_loss contribution value by the maximum value, the result is Output as the highest classification probability result, a boundary frame is generated at the same time, and the type prediction is carried out on the target in the boundary frame;
setting an ECA attention module:
the ECA module adopts a quick one-dimensional convolution mode to replace a full-connection layer of an original structure, and nonlinear characteristic information generated among cross channels is obtained;
setting one-dimensional convolution, wherein the size of a convolution kernel is k, and the convolution kernel represents the coverage range of cross-channel information, namely, the current channel and k adjacent channels participate in the attention of a prediction channel;
k and the dimension C of the total channel have a mapping relation, and when the dimension C of the total channel is determined, a one-dimensional convolution kernel k is obtained through self-adaptive calculation;
projecting the two-dimensional convolution kernel on a space domain to obtain a nonlinear function related to the distribution rule of the image characteristic points, and rapidly determining an optimal solution by using the function, wherein the mapping relation is linear mapping, and the formula is as follows:
Φ(C)＝ak-b，
however, the linear mapping relation is too single, and the number of channels of the convolution network is generally set to be the power of 2, so that the linear function is generalized to the nonlinear function, and the calculation formula is as follows:
C＝Φ(K)＝2 (ak-b)
from the above, the number of channels C is specified, resulting in the formula:
in the above formula: |x| odd For an odd number nearest to x, b=1, a=2;
improving a second layer of Backbone of the wheat scab spore recognition model, and adding an ECA module at the tail end of a CSP residual block of the second layer of Backbone;
setting an ASFF feature fusion module:
the ASFF feature fusion module performs self-adaptive learning on space weight parameters by mapping and fusing the scale features of each layer to obtain a new weighting mode, so as to realize fusion of the features of different layers;
in the three-layer structure of ASFF, level is set 1 、Level 2 、Level 3 Three feature layers of the PANet module are respectively, and three feature layers are fused, wherein the three feature layers are Level 1 、Level 2 、Level 3 The results of (2) and (3) output the features of the three feature layers by X (1), X (2) and X (3) and perform convolution calculation:
multiplying X (1), X (2) and X (3) by weight parameters alpha (3), beta (3) and gamma (3) respectively and summing to obtain feature output ASFF after feature fusion 3 This process formula is expressed as follows:
wherein,,representing a new profile obtained by ASFF, < >>Respectively representing the weight parameters of the three feature layers, which are satisfied by the Softmax function> Features of layers 1, 2, and 3, respectively;
ensuring Level by upsampling or downsampling 1 ，Level 2 ，Level 3 The output characteristic structure of each layer is the same, and the number of channels is kept unchanged;
wherein the lowest ASFF 3 In the process, the Level is firstly calculated 1 、Level 2 The number of channels is compressed by 1×1 convolution to be compared with the Level 3 The same, up-sampling is performed at four times and twice to obtain Level 3 Finally, performing accumulation operation on the same dimension of the number of the blocks;
and the output end of the second layer of the Neck layer FPN+PAN structure of the wheat scab spore recognition model is combined with an ASFF feature fusion module to perform self-adaptive learning on space weight parameters by mapping and fusing scale features of all layers, so that a new weighting mode is obtained, and fusion of features of different layers is realized.
The training of the wheat scab spore recognition model comprises the following steps:
building a Pytorch neural network training environment of Python=3.8 and CUDA=11.6 versions;
setting the input size of an image to 640 multiplied by 640, the confidence threshold value to be 0.5, the initial learning rate to be 0.001, the weight attenuation coefficient to be 0.0005, the size of a model training Batch to be 16, and the training iteration period to be 100;
inputting the spore image dataset into a wheat scab spore recognition model, and training and generating a weight file;
the Input end of the first layer performs Mosaic on data in a mode of mosaics, random scaling, random cutting and random arrangement, scales the image data to 640 x 640 standard size and sends the image data into a detection network;
firstly, slicing the picture by using a Focus of the second backbone layer of the backhaul, and sending W, H characteristic information into a channel space; the CSP module expands the input channel and performs convolution calculation on the input channel to realize feature extraction; the ECA module calculates the channel number C of the CSP features through nonlinear function mapping, and a new convolution kernel k can be obtained from the channel number, so that the CSP features can be adaptively and optimally adjusted. Obtaining a double down sampling characteristic diagram without information loss; the final SPP module converts the multi-size feature images output by the ECA into the feature images and feature vectors with fixed sizes required by the Neck layer through pyramid pooling;
the third layer Neck firstly carries out up-sampling on images with the resolution ratios of 76 multiplied by 76, 38 multiplied by 38 and 19 multiplied by 19 from high to low through an FPN structure to obtain semantic information of target spores; then, the PAN structure is utilized to downsample the resolution ratios from low to high according to 19×19, 38×38 and 76×76, and coordinate information of the target spores is obtained; finally, the characteristic information with the resolution of 38 multiplied by 38 and 76 multiplied by 76 is compressed into the same channel number as 19 multiplied by 19 through convolution calculation by an ASFF module, and then up-sampling is carried out on the characteristic information to ensure that three layers of outputs are positioned at the same latitude, and the obtained weight parameter is used as the characteristic coefficient output, so that the high and low resolutions are fused, and the characteristic information on various scales is fully utilized;
and the fourth layer of Output uses GIOU_Loss as a Loss function of the binding box and calculates the precision. And calculating the precision after each five training steps are completed and generating a weight file. After one hundred times of iterative training is completed, generating average precision and various evaluation indexes, and reserving an optimal weight file in all weight files for detecting and identifying scab spores;
and the generation weight file is utilized to realize rapid and accurate detection and identification of the scab spores.
Advantageous effects
Compared with the prior art, the wheat scab spore identification method based on Yolov5-ECA-ASFF has the advantages that an ECA space attention mechanism is added at the tail end of a CSPNet residual block of a Yolov5s backbone network to strengthen the channel characteristics of an input end characteristic diagram; an ASFF module with a self-adaptive feature fusion mechanism is introduced at the end of a Neck feature extraction network, so that a Yolov5-ECA-ASFF network is constructed by fully utilizing high-level information and bottom-layer features of an input image, and quick and accurate detection and identification of wheat scab spores are effectively realized.
According to the invention, classical single-double-stage target recognition networks Faster-RCNN, yolov4 and Yolov5 are selected to carry out a comparison experiment on a prepared wheat scab microspore data set. The result shows that the network model provided by the invention has the advantages that under the condition that the detection speed is basically unchanged, the identification accuracy P (Precision), recall rate R (Recall), mAP (mean Average Precision) and F1-score values of the Fusarium head blight of wheat respectively reach 98.57%, 96.5%, 98.4% and 97.4%, and the overall evaluation parameters are higher than that of a plurality of main stream target identification networks.
Drawings
FIG. 1 is a process sequence diagram of the present invention;
FIG. 2 is a diagram of a prior art wheat scab spore culture medium;
FIG. 3 is a prior art wheat scab subgraph;
FIG. 4 is a graph showing enhancement of wheat scab spore data according to the present invention;
FIG. 5 is a diagram of a Yolov5s network architecture according to the present invention;
FIG. 6 is a block diagram of the attention mechanism of the ECA in accordance with the present invention;
FIG. 7 is a block diagram of an ASFF feature fusion module according to the present invention;
FIG. 8 is a diagram of a Yolov5-ECA-ASFF network architecture in accordance with the present invention;
FIG. 9 is a diagram of the detection of the original Yolov5 network;
FIG. 10 is a diagram of a detection of a dual-stage network Faster-RCNN network.
Detailed Description
For a further understanding and appreciation of the structural features and advantages achieved by the present invention, the following description is provided in connection with the accompanying drawings, which are presently preferred embodiments and are incorporated in the accompanying drawings, in which:
as shown in FIG. 1, the wheat scab spore identification method based on Yolov5-ECA-ASFF, provided by the invention, comprises the following steps:
first, establishing a spore image data set: and collecting scab spore images, preprocessing the images and enhancing the data to construct a spore image data set.
In practical application, first, wheat scab fungus spore culture is performed: the strain is placed on a solid culture medium for culture, then placed in a constant temperature incubator, and after a certain growth time, the strain is respectively prepared into spore suspension and the solid culture medium containing spores for shooting fungus micrographs, and the fungus micrographs are shown in figure 2. After screening forty-six colonies, eight kinds of fungus spores including fusarium graminearum which mainly causes wheat scab and four kinds of common fungus spores are selected for shooting, wherein the four kinds of fusarium graminearum, fusarium moniliforme, fusarium tax top and the like which cause wheat scab are respectively introduced, and meanwhile, field common spores such as anthracnose, orange colloid spore, hirsutella sinensis and the like which have a certain similarity degree with the wheat scab spores in shape, color and size are respectively introduced for improving the complexity of data sets. Secondly, obtaining a wheat scab spore microscopic image: the diluted spore suspension is sucked out by a thin-mouth dropper and is dripped in the center of a glass slide, a new dropper is used for gently scraping a solid culture medium and mixing the solid culture medium into liquid of the glass slide, and 500 glass slides containing a plurality of spores are manufactured. And placing the prepared glass slides on a microscope stage, randomly selecting 20 fields containing various fungal spores for each glass slide, shooting by using a microscope, selecting a lens of a low power mirror to find a proper target area, focusing and exposing the proper target area by using a high power mirror with the magnification of 10 multiplied by 40, and collecting wheat scab fungal spore microscopic images by using an image collecting system after a stable picture of the system is clear, wherein the image is shown in fig. 3. The experiment acquires microscopic images of 3 spores of pure scab spores, mixed fungus spores, different illumination conditions and the like. A total of 10010 spore microscopic images were collected and the scab spore dataset was evaluated and taken care by plant protection specialists to ensure its authenticity and effectiveness. And most images contain multi-class spores and have higher background complexity, so that the proposed method has high generalization performance and robustness, and the microscope equipment is a Zeiss Axio Vert.A1 laboratory for data acquisition by using an inverted microscope.
The establishment of the spore image dataset comprises the following steps:
(1) From the collected images, 10000 spore image data containing 4 kinds of wheat scab-causing spores and 4 kinds of mixed heterobacteria spores were selected as the target recognition data set.
(2) The total 10000 data sets are enlarged to 20000 by data enhancement, and the enhancement effect is shown in figure 4.
(3) And manually labeling the amplified picture data by using Labelme software, and converting the obtained data set with json file label type into the label file type required by Yolov5 after labeling to construct a spore image data set.
Secondly, constructing a wheat scab spore recognition model: based on the Yolov5 network, an ECA module and an ASFF module are fused to construct a wheat scab spore recognition model. The method comprises the following specific steps:
(1) Based on the Yolov5, a wheat scab spore recognition model is constructed, the structure of the Yolov5 model is shown in fig. 5, and the wheat scab spore recognition model is set to four layers:
the first layer is Input, the Input is used as an Input end, the image can be scaled in a self-adaptive manner, a K-means genetic algorithm is integrated, an optimal anchor frame value of a data set is calculated in a self-adaptive manner, and the small target detection capability is enhanced;
the second layer is a Backbone structure of the Backbone network, which consists of a Focus, a CSP module, an ECA module and an SPP module. Focus integrates information such as width, height, etc. of the input photo in the channel. When the information is not lost, the calculated amount is only 0.44 times of that of the common convolution, and the ECA structure is shown in figure 6. Further, by performing feature mapping on the base layer, the CSP structure is divided into two parts, and then the two parts are fused by adopting a cross-stage hierarchical structure, so that the problem that gradient information repeatedly appears after the main part network of the convolutional neural network is optimized is solved. The ECA module calculates the channel number C of the CSP features through nonlinear function mapping, and a new convolution kernel k can be obtained from the channel number, so that the CSP features can be adaptively and optimally adjusted. Obtaining a double down sampling characteristic diagram without information loss; the SPP module at the tail end is used for determining whether three structures of a convolution layer, a pooling layer and a selection filter need to be applied or not, converting a multi-size feature image output by ECA into a fixed-size feature image and a feature vector needed by a Neck layer, and enhancing the robustness of the network structure in the aspects of spatial layout, changeable detection target forms and the like.
The third layer is a negk structure consisting of up-down sampling structures of fpn+pan+asff. The FPN structure carries out up-sampling on the images with the resolutions of 76 multiplied by 76, 38 multiplied by 38 and 19 multiplied by 19 from high to low to obtain semantic information of the target spores; then, the PAN structure is utilized to downsample the resolution ratios from low to high according to 19×19, 38×38 and 76×76, and coordinate information of the target spores is obtained; finally, the characteristic information with the resolution of 38 multiplied by 38 and 76 multiplied by 76 is compressed into the same channel number as 19 multiplied by 19 through convolution calculation by an ASFF module, then up-sampling is carried out on the characteristic information to enable three layers of output to be in the same latitude, the obtained weight parameters are used as characteristic coefficient output, the fusion of high resolution and low resolution is realized, an ASFF structure is shown in fig. 7, characteristic information on various scales is fully utilized, and characteristics extracted by different detection layers are fused by each trunk layer to obtain more effective information which is transmitted to a prediction layer.
The fourth layer is Output, the Output end is used for evaluating whether the detection and positioning of the algorithm on the target are accurate or not, the accuracy is calculated by using a GIOU Loss function, the GIOU Loss function eliminates the residual prediction result of dividing the GIOU_Loss contribution value by the maximum value, the result is Output as the highest classification probability result, a boundary frame is generated at the same time, and the type prediction is carried out on the target in the boundary frame.
(2) Setting an ECA attention module:
the ECA module adopts a quick one-dimensional convolution mode to replace a full-connection layer of an original structure, and nonlinear characteristic information generated among cross channels is obtained;
setting one-dimensional convolution, wherein the size of a convolution kernel is k, and the convolution kernel represents the coverage range of cross-channel information, namely, the current channel and k adjacent channels participate in the attention of a prediction channel;
k and the dimension C of the total channel have a mapping relation, and when the dimension C of the total channel is determined, a one-dimensional convolution kernel k is obtained through self-adaptive calculation;
projecting the two-dimensional convolution kernel on a space domain to obtain a nonlinear function related to the distribution rule of the image characteristic points, and rapidly determining an optimal solution by using the function, wherein the mapping relation is linear mapping, and the formula is as follows:
Φ(C)＝ak-b，
however, the linear mapping relation is too single, and the number of channels of the convolution network is generally set to be the power of 2, so that the linear function is generalized to the nonlinear function, and the calculation formula is as follows:
C＝Φ(K)＝2 (ak-b)
from the above, the specified channel number C, the formula can be obtained:
in the above formula: |x| odd For an odd number nearest to x, b=1, a=2.
(3) The second layer Backbone of the wheat scab spore recognition model was modified and an ECA module was added at the end of its CSP residual block.
The second-layer backbox of the wheat scab spore recognition model is improved, the Focus, CSP and SPP structures are simple cross-level fusion for feature extraction of small targets, and detailed features are not concerned, so that an ECA module is added at the tail end of a CSPNet residual block of a Backbone network for enhancing channel features by inputting feature graphs, self-adaptive optimal weight adjustment is carried out on the detailed features, and feature extraction capability of the network for different channels is improved.
(4) Setting an ASFF feature fusion module:
the ASFF feature fusion module performs self-adaptive learning on space weight parameters by mapping and fusing the scale features of each layer to obtain a new weighting mode, so as to realize fusion of the features of different layers;
in the three-layer structure of ASFF, level is set 1 、Level 2 、Level 3 Three feature layers of the PANet module are respectively, and three feature layers are fused, wherein the three feature layers are Level 1 、Level 2 、Level 3 The results of (2) and (3) output the features of the three feature layers by X (1), X (2) and X (3) and perform convolution calculation:
multiplying X (1), X (2) and X (3) by weight parameters alpha (3), beta (3) and gamma (3) respectively and summing to obtain feature output ASFF after feature fusion 3 This process formula is expressed as follows:
wherein,,representing a new profile obtained by ASFF, < >>Respectively representing the weight parameters of the three feature layers, which are satisfied by the Softmax function> Features of layers 1, 2, and 3, respectively;
ensuring Level by upsampling or downsampling 1 ，Level 2 ，Level 3 The output characteristic structure of each layer is the same, and the number of channels is kept unchanged;
wherein the lowest ASFF 3 In the process, the Level is firstly calculated 1 、Level 2 The number of channels is compressed by 1×1 convolution to be compared with the Level 3 Identical, respectively atFour times, up-sampling at twice to obtain Level 3 And finally, performing accumulation operation.
(5) And the output end of the second layer of the Neck layer FPN+PAN structure of the wheat scab spore recognition model is combined with an ASFF feature fusion module to perform self-adaptive learning on space weight parameters by mapping and fusing scale features of all layers, so that a new weighting mode is obtained, and fusion of features of different layers is realized.
The invention improves the Neck of YOLOv5s, realizes small-target multi-layer feature output through the FPN+PAN structure, but the method simply converts the feature images into the same size and then accumulates, and the method can not fully utilize the characteristics on various scales, so that the invention performs self-adaptive learning on space weight parameters by mapping and fusing the scale features of various layers at the output end of the Neck layer FPN+PAN structure in combination with an ASFF module, thus obtaining a new weighting mode, realizing the fusion of the features of different layers, and realizing the full utilization of the high-layer information and the bottom layer features of the image by a network by virtue of a new network structure shown in FIG. 8.
Thirdly, training a wheat scab spore recognition model: the spore image dataset was input into a wheat scab spore recognition model for training. The training of the wheat scab spore recognition model comprises the following steps:
(1) A Pytorch neural network training environment of python=3.8 and cuda=11.6 version was constructed.
(2) The input size of the image is set to 640 multiplied by 640, the confidence threshold is set to 0.5, the initial learning rate is set to 0.001, the weight attenuation coefficient is set to 0.0005, the size of the model training Batch is set to 16, and the training iteration period Epoch is set to 100.
(3) And inputting the spore image dataset into a wheat scab spore recognition model, and completing training and generating an optimal weight file.
A1 The Input end of the first layer performs Mosaic on the data in a mode of mosaics, random scaling, random cutting and random arrangement, then scales the image data to 640 x 640 standard size, and then sends the image data into a detection network;
a2 Firstly, slicing the picture by using a Focus of the second backbone layer of the backhaul, and sending W, H characteristic information into a channel space; the CSP module expands the input channel and performs convolution calculation on the input channel to realize feature extraction; the ECA module calculates the channel number C of the CSP features through nonlinear function mapping, and a new convolution kernel k can be obtained from the channel number, so that the CSP features can be adaptively and optimally adjusted. Obtaining a double down sampling characteristic diagram without information loss; the final SPP module converts the multi-size feature images output by the ECA into the feature images and feature vectors with fixed sizes required by the Neck layer through pyramid pooling;
a3 The third layer Neck firstly carries out up-sampling on images with the resolution ratios of 76 multiplied by 76, 38 multiplied by 38 and 19 multiplied by 19 from high to low through an FPN structure to obtain semantic information of target spores; then, the PAN structure is utilized to downsample the resolution ratios from low to high according to 19×19, 38×38 and 76×76, and coordinate information of the target spores is obtained; finally, the characteristic information with the resolution of 38 multiplied by 38 and 76 multiplied by 76 is compressed into the same channel number as 19 multiplied by 19 through convolution calculation by an ASFF module, and then up-sampling is carried out on the characteristic information to ensure that three layers of outputs are positioned at the same latitude, and the obtained weight parameter is used as the characteristic coefficient output, so that the high and low resolutions are fused, and the characteristic information on various scales is fully utilized;
a4 Fourth layer Output adopts GIOU_Loss as the Loss function of the binding box and calculates the precision. And calculating the precision after each five training steps are completed and generating a weight file. After one hundred times of iterative training is completed, generating average precision and various evaluation indexes, and reserving an optimal weight file in all weight files for detecting and identifying scab spores;
(4) And the generation weight file is utilized to realize rapid and accurate detection and identification of the scab spores.
Fourth, obtaining wheat scab spores to be identified: and obtaining wheat scab spores to be identified, and performing pretreatment.
Fifthly, obtaining a wheat scab spore identification result to be identified: and inputting the pretreated wheat scab spores to be identified into a trained wheat scab spore identification model to obtain a wheat scab spore identification result.
In order to verify the accuracy of wheat scab spore detection and identification, 2001 complex wheat scab spore images in a test set are tested by using the obtained weight file.
The invention takes the Precision P (Precision), F1-score, recall R (Recall) and mAP50 (mean Average Precision) of each model as the evaluation index of the experiment. P and R represent the accuracy of the detection algorithm in the positive and correct samples, respectively. However, P and R cannot directly evaluate the detection accuracy. Thus, mAP50 and F1 indices were used to evaluate the ability of the detection algorithm. Accuracy refers to the likelihood that the detected positive sample is actually still a positive sample, and recall refers to the likelihood that the actual positive sample is found; mAP50 is the average AP value at which all class detection results intersect the predicted and real boxes, and the ratio IoU (Intersection Over Union) is thresholded to 0.5. The higher the mAP50 and F1 indexes are, the higher the network precision is, and the value range of the six evaluation indexes is 0-1 (the smaller the numerical value is, the worse the segmentation effect is). The method for calculating the six indexes comprises the following steps:
where A, B is the size of the prediction box and the real box, respectively. TP (True Positives) is the number of actual detected objects in the data set. FP (False Positives) refers to the number of false detection objects of the detection model. FN (False Negatives) refers to the number of objects in the detection model that are missed.
Meanwhile, in order to verify the effectiveness of the detection algorithm provided by the invention, two characteristic fusion mechanisms of ASFF and BiFPN are respectively added into a backlight of Yolov5s in the research. The results of the comparative experiments are shown in Table 1.
Table 1 results of comparative experiments with different feature fusion mechanism modules
The result shows that the two feature fusion mechanisms are improved on the detection result, the ASFF adaptively learns the spatial weight of 3-layer feature mAP fusion after the same scaling operation, and compared with the traditional cascading type multi-level feature fusion method, the method has good superiority, and compared with the Yolov5s, the mAP is improved by 2.8%. The BiFPN adopts a weighted bidirectional feature pyramid mode, firstly realizes high-efficiency bidirectional cross-scale linking and then realizes weighted feature graph fusion, so that mAP is improved by 2.3% compared with Yolov5s, but because the bi-directionality of the BiFPN is compared with that of PANet, only one node is deleted, the calculated amount of the BiFPN is large, and the detection speed is reduced by 6 frames compared with ASFF. Therefore, according to the comparison experiment result, the network model combined with the ASFF attention fusion mechanism is optimal in each performance
And then respectively selecting a channel attention mechanism ECA, a mixed domain attention mechanism CA and a CBAM to carry out an ablation experiment. A comparison experiment was developed and the results of which were shown in Table 2 were obtained by constructing and comparing Yolov5-CA-ASFF, yolov5-ECA-ASFF and Yolov5-CBAM-ASFF, respectively. The results show that the three attention mechanisms improve the characteristic detection capability of the fungal spores, the calculated amount is improved, the time consumption is increased, and the detection rate is reduced. In terms of accuracy performance, ECA and CBAM differ by only one percentage point, but because the ECA module avoids the strategy of partial cross-channel interaction, the method of self-adaptive one-dimensional convolution kernel is used for determining the coverage rate of partial cross-channel interaction, and the detection rate is obviously superior to that of CBAM. Thus, the ECA attention mechanism introduced in the paper is more suitable for improving the detection capability of wheat scab spores inserted into a network.
Table 2 different attentiveness mechanisms module vs. experimental results
After the two types of comparison experiments, a network model Yolov5-ECA-ASFF aiming at fungus spore detection is obtained, but the robustness of the network is not verified, so that in order to further verify that the Yolov5-ECA-ASFF is used for carrying out robustness test and performance comparison on the network by utilizing the two data under the condition of multi-fungus environment and dust particle interference, a data set which is prepared by introducing mixed shooting of mixed fungus spores with a certain similarity degree of morphology, color and size with wheat scab spores is selected. The mixed dataset was compared with Yolov4 and Yolov5s using the modified network structure, respectively, and the results of the measurements are shown in table 3. The comparison results of the detection of the normal illumination data and the low illumination data are shown in fig. 9 and fig. 10 respectively. Under the condition of higher complexity of the data set, compared with other methods, the improved network still has obvious improvement under each detection index, and the spore detection performance still has certain superiority.
Table 3 results of comparative experiments with Yolov4, yolov5-ECA-ASFF in two different scenarios, dataset A (purebred gibberellic disease spores) and dataset B (Mixed spores)
As the wheat scab fungus spore target is smaller, the invention provides a structural network of Yolov5-ECA-ASFF, which is used for strengthening the condition of poor detection of the small target. In order to evaluate the performance of the detection model in the mainstream network, the present section performs network training under the same model training environment and parameter configuration with the proposed network structure, a single-order YOLO series network and a representative double-order fast-RCNN target detection network, and performs a comparison experiment under the same data set by using four network models of Yolov4, yolov5s, yolov5-ECA-ASFF and fast-RCNN. Wherein the metrics include precision, recall, AP value, F1 score, weight parameters, and detection rate. The modified network Yolov5-ECA-ASFF was mainly compared with Yolov4 and Yolov5 s. The results of the wheat scab spore detection for the various networks described above are shown in table 4.
Table 4 comparison experiment results table of classical single-double-order network and improved network
As can be seen from the results in table 4, compared with other methods, the improved network structure ensures the detection rate and simultaneously has significantly improved mAP values and F1 score. When the same data set is used for training and testing, compared with Yolov4 and Yolov5s, P is respectively improved by 5.9% and 6.7%, R is respectively improved by 4.4% and 4.0%, and mAP is respectively improved by 4.8% and 5.6%; the model parameter and the detection rate are both superior to those of Yolov4, and the detection algorithm only reduces the detection rate of 1 frame under the condition of increasing the model parameter of 1.8MB, so that the wheat scab spore detection model provided by the invention is reliable. Compared with the original Yolov5s algorithm, the algorithm in the invention obtains better performance in mAP values in different IoU threshold value intervals.
The foregoing has shown and described the basic principles, principal features and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, and that the above embodiments and descriptions are merely illustrative of the principles of the present invention, and various changes and modifications may be made therein without departing from the spirit and scope of the invention, which is defined by the appended claims. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (4)
1. A wheat scab spore identification method based on Yolov5-ECA-ASFF is characterized by comprising the following steps:
11 Establishment of spore image dataset: collecting and collecting scab spore images, preprocessing the images and enhancing the data to construct a spore image data set;
12 Building a wheat scab spore recognition model: based on a Yolov5 network, an ECA module and an ASFF module are fused to construct a wheat scab spore recognition model;
13 Training of wheat scab spore recognition model: inputting the spore image dataset into a wheat scab spore recognition model for training;
14 Obtaining wheat scab spores to be identified: acquiring wheat scab spores to be identified, and preprocessing;
15 Obtaining the identification result of wheat scab spores to be identified: and inputting the pretreated wheat scab spores to be identified into a trained wheat scab spore identification model to obtain a wheat scab spore identification result.
2. The method for identifying wheat scab spores based on Yolov5-ECA-ASFF of claim 1, wherein the establishment of the spore image dataset comprises the steps of:
21 Selecting 10000 spore image data containing 4 kinds of wheat scab and 4 kinds of mixed fungus spore image data of 4 kinds of scab fungus spores from the acquired images as a target identification data set;
22 Expanding the total 10000 data sets to 20000 data sets through data enhancement;
23 Manually labeling the amplified picture data by Labelme software, and converting the obtained label type dataset with json file into the label file type required by Yolov5 after labeling to construct a spore image dataset.
3. The wheat scab spore recognition method based on Yolov5-ECA-ASFF of claim 1, wherein the constructing the wheat scab spore recognition model includes the steps of:
31 Building a wheat scab spore recognition model based on Yolov5, and setting the wheat scab spore recognition model into four layers:
the first layer is Input, the Input is used as an Input end for adaptively scaling an image, a K-means genetic algorithm is integrated, an optimal anchor frame value of a data set is adaptively calculated, and the small target detection capability is enhanced;
the second layer is a Backbone structure of the Backbone network, and the Backbone structure comprises a Focus, a CSP module and an SPP module; the Focus integrates the width and height information of the input photo in the channel, the CSP structure is divided into two parts by carrying out feature mapping on the base layer, a cross-stage hierarchical structure is adopted to fuse the two parts, and an SPP module at the tail end is used for three structures of a convolution layer, a pooling layer and a selection filter;
the third layer is a Neck structure which comprises an up-down sampling structure FPN layer of FPN+PAN, and the tail end is combined with a PAN module to carry out up-sampling;
the fourth layer is Output, the Output end is used for evaluating whether the detection and positioning of the algorithm on the target are accurate or not, the accuracy is calculated by using a GIOU Loss function, the GIOU Loss function eliminates the residual prediction result of dividing the GIOU_loss contribution value by the maximum value, the result is Output as the highest classification probability result, a boundary frame is generated at the same time, and the type prediction is carried out on the target in the boundary frame;
32 Setting ECA attention module:
the ECA module adopts a quick one-dimensional convolution mode to replace a full-connection layer of an original structure, and nonlinear characteristic information generated among cross channels is obtained;
setting one-dimensional convolution, wherein the size of a convolution kernel is k, and the convolution kernel represents the coverage range of cross-channel information, namely, the current channel and k adjacent channels participate in the attention of a prediction channel;
k and the dimension C of the total channel have a mapping relation, and when the dimension C of the total channel is determined, a one-dimensional convolution kernel k is obtained through self-adaptive calculation;
projecting the two-dimensional convolution kernel on a space domain to obtain a nonlinear function related to the distribution rule of the image characteristic points, and rapidly determining an optimal solution by using the function, wherein the mapping relation is linear mapping, and the formula is as follows:
Φ(C)＝ak-b，
however, the linear mapping relation is too single, and the number of channels of the convolution network is generally set to be the power of 2, so that the linear function is generalized to the nonlinear function, and the calculation formula is as follows:
C＝Φ(K)＝2 (ak-b)
from the above, the number of channels C is specified, resulting in the formula:
in the above formula: x is x odd For an odd number nearest to x, b=1, a=2;
33 Improving a second layer of Backbone of the wheat scab spore recognition model, and adding an ECA module at the tail end of a CSP residual block of the second layer of Backbone;
34 Setting ASFF feature fusion module):
the ASFF feature fusion module performs self-adaptive learning on space weight parameters by mapping and fusing the scale features of each layer to obtain a new weighting mode, so as to realize fusion of the features of different layers;
in the three-layer structure of ASFF, level is set 1 、Level 2 、Level 3 Three feature layers of the PANet module are respectively, and three feature layers are fused, wherein the three feature layers are Level 1 、Level 2 、Level 3 The results of (2) and (3) output the features of the three feature layers by X (1), X (2) and X (3) and perform convolution calculation:
multiplying X (1), X (2) and X (3) by weight parameters alpha (3), beta (3) and gamma (3) respectively and summing to obtain feature output ASFF after feature fusion 3 This process formula is expressed as follows:
wherein,,representing a new profile obtained by ASFF, < >>Respectively representing the weight parameters of the three feature layers, which are satisfied by the Softmax function> Features of layers 1, 2, and 3, respectively;
ensuring Level by upsampling or downsampling 1 ，Level 2 ，Level 3 The output characteristic structure of each layer is the same, and the number of channels is kept unchanged;
wherein the lowest ASFF 3 In the process, the Level is firstly calculated 1 、Level 2 The number of channels is compressed by 1×1 convolution to be compared with the Level 3 The same, up-sampling is performed at four times and twice to obtain Level 3 Finally, performing accumulation operation on the same dimension of the number of the blocks;
35 The output end of the second layer of the Neck layer FPN+PAN structure of the wheat scab spore recognition model is combined with an ASFF feature fusion module to perform adaptive learning on space weight parameters by mapping and fusing scale features of all layers, so that a new weighting mode is obtained, and fusion of features of different layers is realized.
4. The method for identifying wheat scab spores based on Yolov5-ECA-ASFF according to claim 1, wherein the training of the wheat scab spore identification model comprises the steps of:
41 Pytorch neural network training environment of version python=3.8 and cuda=11.6 is built;
42 Setting the input size of the image to 640 multiplied by 640, the confidence threshold value to 0.5, the initial learning rate to 0.001, the weight attenuation coefficient to 0.0005, the size of the model training Batch to 16, and the training iteration period to 100;
43 Inputting the spore image dataset into a wheat scab spore recognition model, and completing training and generating a weight file;
431 The Input end of the first layer performs Mosaic on the data in a mode of mosaics, random scaling, random cutting and random arrangement, then scales the image data to 640 x 640 standard size, and then sends the image data into a detection network;
432 Firstly, slicing the picture by using a Focus of the second backbone layer of the backhaul, and sending W, H characteristic information into a channel space; the CSP module expands the input channel and performs convolution calculation on the input channel to realize feature extraction; the ECA module calculates the channel number C through nonlinear function mapping on the CSP features, a new convolution kernel k can be obtained from the channel number, self-adaptive optimal weight adjustment of the CSP features is realized, and a double downsampling feature map without information loss is obtained; the final SPP module converts the multi-size feature images output by the ECA into the feature images and feature vectors with fixed sizes required by the Neck layer through pyramid pooling;
433 The third layer Neck firstly carries out up-sampling on images with the resolution ratios of 76 multiplied by 76, 38 multiplied by 38 and 19 multiplied by 19 from high to low through an FPN structure to obtain semantic information of target spores; then, the PAN structure is utilized to downsample the resolution ratios from low to high according to 19×19, 38×38 and 76×76, and coordinate information of the target spores is obtained; finally, the characteristic information with the resolution of 38 multiplied by 38 and 76 multiplied by 76 is compressed into the same channel number as 19 multiplied by 19 through convolution calculation by an ASFF module, and then up-sampling is carried out on the characteristic information to ensure that three layers of outputs are positioned at the same latitude, and the obtained weight parameter is used as the characteristic coefficient output, so that the high and low resolutions are fused, and the characteristic information on various scales is fully utilized;
434 Fourth layer Output adopts GIOU_Loss as a Loss function of the binding box and calculates the precision, and calculates the precision and generates a weight file after the training is completed every five times; after one hundred times of iterative training is completed, generating average precision and various evaluation indexes, and reserving an optimal weight file in all weight files for detecting and identifying scab spores;
44 And (3) quick and accurate detection and identification of the scab spores are realized by using the generation weight file.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310444199.6A CN116524255A (en) | 2023-04-24 | 2023-04-24 | Wheat scab spore identification method based on Yolov5-ECA-ASFF |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310444199.6A CN116524255A (en) | 2023-04-24 | 2023-04-24 | Wheat scab spore identification method based on Yolov5-ECA-ASFF |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116524255A true CN116524255A (en) | 2023-08-01 |
Family
ID=87393473
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310444199.6A Pending CN116524255A (en) | 2023-04-24 | 2023-04-24 | Wheat scab spore identification method based on Yolov5-ECA-ASFF |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116524255A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116821698A (en) * | 2023-08-31 | 2023-09-29 | 中国科学技术大学 | Wheat scab spore detection method based on semi-supervised learning |
-
2023
- 2023-04-24 CN CN202310444199.6A patent/CN116524255A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116821698A (en) * | 2023-08-31 | 2023-09-29 | 中国科学技术大学 | Wheat scab spore detection method based on semi-supervised learning |
| CN116821698B (en) * | 2023-08-31 | 2024-01-05 | 中国科学技术大学 | Wheat scab spore detection method based on semi-supervised learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112396002B (en) | SE-YOLOv 3-based lightweight remote sensing target detection method | |
| CN109614985B (en) | Target detection method based on densely connected feature pyramid network | |
| CN107609601B (en) | Ship target identification method based on multilayer convolutional neural network | |
| CN110956185B (en) | Method for detecting image salient object | |
| CN109584248B (en) | Infrared target instance segmentation method based on feature fusion and dense connection network | |
| CN112418117B (en) | Small target detection method based on unmanned aerial vehicle image | |
| CN111310861A (en) | License plate recognition and positioning method based on deep neural network | |
| Saito et al. | Building and road detection from large aerial imagery | |
| CN111368896A (en) | Hyperspectral remote sensing image classification method based on dense residual three-dimensional convolutional neural network | |
| CN112488210A (en) | Three-dimensional point cloud automatic classification method based on graph convolution neural network | |
| KR102138657B1 (en) | Apparatus and method for robust face recognition via hierarchical collaborative representation | |
| CN111583263A (en) | Point cloud segmentation method based on joint dynamic graph convolution | |
| CN112347970B (en) | Remote sensing image ground object identification method based on graph convolution neural network | |
| CN108052966A (en) | Remote sensing images scene based on convolutional neural networks automatically extracts and sorting technique | |
| CN110222767A (en) | Three-dimensional point cloud classification method based on nested neural and grating map | |
| CN113486764A (en) | Pothole detection method based on improved YOLOv3 | |
| CN112733614A (en) | Pest image detection method with similar size enhanced identification | |
| CN115116054A (en) | Insect pest identification method based on multi-scale lightweight network | |
| CN116524255A (en) | Wheat scab spore identification method based on Yolov5-ECA-ASFF | |
| CN113657414B (en) | Object identification method | |
| CN117542082A (en) | Pedestrian detection method based on YOLOv7 | |
| CN109102019A (en) | Image classification method based on HP-Net convolutional neural networks | |
| CN113989612A (en) | Remote sensing image target detection method based on attention and generation countermeasure network | |
| CN115761240B (en) | Image semantic segmentation method and device for chaotic back propagation graph neural network | |
| Sosa-Trejo et al. | Vision-based techniques for automatic marine plankton classification |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |