Search

CN-122023886-A - Intelligent identification method for citrus diseases

CN122023886ACN 122023886 ACN122023886 ACN 122023886ACN-122023886-A

Abstract

The invention discloses an intelligent identification method for citrus diseases, which relates to the technical field of image processing and computer vision, and comprises the steps of firstly obtaining an image of citrus leaves, removing high-frequency noise by Gaussian filtering, extracting the leaf profile by a Canny operator, generating a binary mask, and removing a background area by mask operation; the method comprises the steps of constructing a first model branch based on DenseNet121 and a second model branch based on EfficientNetB7, removing a top classifier, accessing to a global averaging pooling layer, training models by using a dynamic learning rate until convergence, and fusing probability vectors of dual-model output by using a weighted averaging method in a prediction stage, wherein the weight coefficients of the first model and the second model are 0.75 and 0.25 respectively. According to the invention, through cloud edge collaborative architecture implementation, complex background interference is effectively eliminated by combining a specific preprocessing denoising and dual-model weighting integration strategy, a fuzzy judgment interval is eliminated by utilizing the characteristic complementation advantage, and the identification precision and reliability of citrus diseases in a mobile terminal application scene are remarkably improved.

Inventors

  • PENG DONGHAI

Assignees

  • 韶关学院

Dates

Publication Date
20260512
Application Date
20260113

Claims (7)

  1. 1. An intelligent identification method for citrus diseases is characterized by comprising the following steps: S1, obtaining and preprocessing image data, namely obtaining an image dataset of the citrus blade, wherein the image dataset of the citrus blade comprises healthy blades and images with specific disease categories, carrying out Gaussian filter smoothing processing on the images to remove high-frequency noise, carrying out edge detection on the images subjected to the Gaussian filter smoothing processing by adopting a Canny operator to extract blade profile characteristics, generating a binarization mask based on the extracted blade profile characteristics, carrying out bit-wise and operation on the binarization mask and the images subjected to the Gaussian filter smoothing processing, reserving a blade area and removing a background area; Step S2, constructing a dual-channel transfer learning model, namely constructing a first model branch and a second model branch which work in parallel, wherein the first model branch is based on DenseNet and the second model branch is based on EfficientNetB and the first model branch is based on a network structure of EfficientNetB; Step S3, model training, namely dividing the data set preprocessed in the step S1 into a training set and a verification set, and training the first model branch and the second model branch by utilizing a dynamic learning rate scheduling function until the first model branch and the second model branch converge; s4, an integration decision based on weighted average is that an orange leaf image to be identified is obtained, and is preprocessed in the step S1 and then is input into a first model branch and a second model branch after training is completed; The first model branch outputs a first probability vector, and the second model branch outputs a second probability vector; fusing the first probability vector and the second probability vector by adopting a weighted average method, calculating a final classification probability vector, and determining a recognition result according to the final classification probability vector; the specific calculation mode of the weighted average method comprises the steps of multiplying a first probability vector by a first weight coefficient to obtain a first weighted vector, multiplying a second probability vector by a second weight coefficient to obtain a second weighted vector, and adding the first weighted vector and the second weighted vector to obtain a final classification probability vector.
  2. 2. The intelligent citrus disease identification method according to claim 1, wherein the specific operations of image data acquisition and preprocessing further comprise: and carrying out data enhancement on the preprocessed image, wherein the data enhancement mode comprises horizontal overturn, vertical overturn and Gaussian noise addition.
  3. 3. The intelligent citrus disease identification method according to claim 1, wherein the specific operations of image denoising and edge extraction are as follows: Performing convolution operation on the original RGB image by using Gaussian kernel to execute Gaussian filtering; setting a high threshold and a low threshold, and calculating the amplitude and the direction of the image gradient by using a Canny operator; the background area pixel value is set to 0.
  4. 4. The intelligent citrus disease identification method according to claim 1, wherein in the step S2, the number of nodes of the full connection layer is set to 4; the specific disease category comprises anthracnose, canker and yellow dragon disease, and the identification result is healthy, anthracnose, canker or yellow dragon disease and the probability percentage corresponding to healthy, anthracnose, canker or yellow dragon disease.
  5. 5. The intelligent citrus disease identification method according to claim 1, wherein in step S3, the specific environment for model training is configured to: The dynamic learning rate scheduling function is configured to set a larger learning rate at an early stage of training, and to decrease the learning rate according to a preset decay strategy as training rounds increase.
  6. 6. The intelligent citrus disease identification method according to claim 1, wherein in the step S4, the first weight coefficient and the second weight coefficient are configured as follows: setting a second weight coefficient corresponding to the second model branch to be 0.25; The first weight coefficient corresponding to the first model branch is set to be 0.75; based on the setting of the first weight coefficient and the second weight coefficient, the number of samples with the prediction probability larger than 0.9 in the final classification probability vector is maximum, and the number of samples with the prediction probability smaller than 0.5 is 0.
  7. 7. The intelligent citrus disease identification method according to claim 1, wherein the intelligent citrus disease identification method is implemented by adopting a cloud edge cooperative architecture, and comprises the following steps: Calling a device camera to collect an image, compressing the image data, sending the compressed image data to a cloud server through an HTTP POST request, and receiving JSON format data returned by the cloud server to display a result; The cloud server side comprises the steps of constructing a Web API interface by adopting a Linux operating system and Flask framework, loading a trained integrated model weight file, receiving an image data stream sent by a mobile terminal, executing the steps S1 to S4, generating JSON format data containing disease category names and corresponding confidence values, and returning.

Description

Intelligent identification method for citrus diseases Technical Field The invention relates to the technical field of image processing and computer vision, in particular to an intelligent citrus disease identification method. Background Citrus is one of the global important commercial crops, and its planting scale and yield are growing year by year. In the growth cycle of citrus, the occurrence of common diseases such as anthracnose, canker, yellow dragon and the like can seriously affect the yield and quality of fruits, and if prevention and control measures cannot be found and taken in time, the spread of the diseases can bring huge economic loss to growers. Therefore, the method can rapidly and accurately identify and diagnose the citrus leaf diseases, and has important significance for guiding the accurate pesticide application of agriculture and guaranteeing the healthy development of industry. Traditional disease identification mainly relies on plant protection experts or manual experience judgment of farmers, and the mode is limited by subjective factors and has low efficiency, so that the inspection requirement of a large-scale orchard is difficult to meet. Along with the development of computer vision and deep learning technology, the convolutional neural network-based image recognition technology is gradually applied to the field of agricultural disease diagnosis. In the prior art, a mobile terminal or acquisition equipment is generally utilized to acquire citrus leaf images in an orchard, the acquired images are directly input into a general deep learning network model for feature extraction and classification, and an attempt is made to replace manual visual inspection by an automatic means so as to improve the disease identification efficiency. However, in actual natural orchard scenarios, the captured images often contain complex background environments, such as soil, weeds, and changing lighting conditions. When processing images in such non-laboratory environments, the existing identification methods generally lack a targeted target region extraction and background rejection mechanism, and directly use original images containing a large amount of background noise for model training or reasoning. When the model is subjected to feature learning, invalid textures in the background or pseudo features generated by uneven illumination are easily extracted, so that the background noise interferes with key pathological features of the blade diseases, the feature extraction capability of the model in a complex field environment is weakened, and the recognition accuracy and stability cannot meet the high-precision requirements of actual agricultural production. Disclosure of Invention Aiming at the defects of the prior art, the invention provides an intelligent identification method for citrus diseases, which solves the problems that the artificial identification efficiency of citrus diseases is low, the identification precision of a single deep learning model is insufficient under a complex orchard background in the prior art, and a high-precision model is difficult to directly deploy and operate on mobile terminal equipment. In order to achieve the purpose, the intelligent identification method for the citrus diseases is realized through the following technical scheme that the intelligent identification method comprises the following steps: Image data acquisition and preprocessing: a citrus leaf image dataset is obtained, the dataset comprising images of healthy leaves and specific disease categories. The image is subjected to gaussian filter smoothing processing to remove high-frequency noise. And carrying out edge detection on the image subjected to Gaussian filter smoothing treatment by adopting a Canny operator, and extracting blade contour features. And generating a binarization mask based on the extracted blade profile features, and performing bit-wise AND operation on the binarization mask and the Gaussian filter smoothed image, so as to reserve a blade area and remove a background area. Constructing a dual-channel transfer learning model: A first model branch and a second model branch are constructed that work in parallel. The first model branch is based on DenseNet network architecture and the second model branch is based on EfficientNetB network architecture. The first model branch and the second model branch are loaded with pre-training weights of an ImageNet data set, the top classifier of the original network is removed, and the top classifier is respectively connected to a global average pooling layer and a full connection layer with an activation function of Softmax. Model training: Dividing the preprocessed data set into a training set and a verification set, and training the first model branch and the second model branch by utilizing a dynamic learning rate scheduling function until the first model branch and the second model branch converge. An integrated decision step based on weig