| The navel orange industry is one of the most iconic characteristic industries in the Gannan area,but it often suffers from diseases and pests during its growth,severely affecting the quality and yield of the fruit.With consumers’ increasing demands for the quality of agricultural products and the improvement of artificial intelligence technology,machine vision technology has been widely applied in the identification of crop diseases and pests,improving the accuracy and timeliness of disease and pest detection.However,as a local cash crop,there has been relatively little research on navel orange disease identification based on machine vision,and the disease images in existing studies are often collected against specific backgrounds,which cannot meet the needs of farmers and agricultural experts in real production environments.Research on navel orange disease image recognition methods can provide technical assurance for the automated,intelligent,and scientific management of navel orange cultivation.There are many large-scale optimization problems in disease image recognition issues that traditional mathematical programming methods are difficult to solve,whereas swarm intelligence optimization algorithms can obtain optimal or sub-optimal solutions in a short time.Combining with the characteristics of navel orange leaf disease image,many intelligent optimization algorithms were studied,and it was found that particle swarm optimization algorithm and cuckoo search algorithm were suitable to deal with many optimization problems in navel orange leaf disease image recognition.Therefore,based on the improved particle swarm optimization algorithm and cuckoo search algorithm,the problem of navel orange leaf disease recognition was studied in this paper.The specific work of the paper is as follows:(1)The classical swarm intelligent optimization algorithm Particle Swarm Optimization algorithm and Cuckoo Search algorithm are studied and improvedIn order to solve the problems of prematurity and slow convergence when PSO is applied to disease image recognition,three improvement strategies are proposed.Firstly,chaos mechanism is introduced to enhance the diversity of population and improve the convergence accuracy.The second is to adopt the clustering mechanism and adaptive inertia weight to accelerate the convergence speed;The third is to add precocious judgement in the later stage of the algorithm.Based on the above three strategies,a Chaotic Group-based Niche PSO(CGNPSO)is proposed,and experiments show that the search efficiency and accuracy of this algorithm have been effectively improved.In order to solve the problem that Cuckoo Search algorithm is sensitive to control parameters and uses random walk to generate new solutions easy vibration when it is applied to disease image processing,two improved strategies are proposed.One is to adapt the inertia weight to improve the efficiency of search;The second is to use dynamic random walk to generate new solutions and improve the convergence accuracy of the algorithm.Based on these two improved strategies,an Improved Cuckoo Search(ICS)algorithm is proposed,and experiments show that the algorithm is effective.(2)A nonlinear transformation image enhancement method for navel orange leaf disease based on swarm intelligence optimization is proposedAiming at the condition that navel orange leaf disease image taken in field environment is easily affected by the environment and the background is relatively complex,the disease image is preprocessed,and the nonlinear transform enhancement method based on swarm intelligence optimization is proposed to enhance the navel orange lear disease image.Because the normalized incomplete Beta function can automatically fit four kinds of image transformation curves,taking into account the global and local features of the image,this paper selects the incomplete Beta function for nonlinear enhancement of the image.To solve the problem that the parameters of the incomplete Beta function are difficult to be determined,two improved swarm intelligence algorithms are proposed in this paper to optimize the parameters.The experimental results show that the nonlinear transformation enhancement method based on CGNPSO has a good enhancement effect on navel orange leaf disease image.(3)A multi-strategy segmentation method for navel orange leaf disease image was proposedA multi-strategy image segmentation method based on ICS and CGNPSO was proposed for navel orange leaf disease image with complex background and difficulty in segmentation.In this method,the disease image is segmented preliminarily based on maximum inter-class variance(OTSU),and then the results of multi-threshold segmentation are used as the input image of fuzzy C-means(FCM)clustering segmentation method for further segmentation.Preliminary multi-threshold segmentation can help FCM find out the initial cluster center,and effectively reduce the gray gradient of the image,and reduce the number of iterations and running time for the subsequent FCM segmentation.The multi-strategy segmentation method proposed in this paper makes full use of swarm intelligence optimization to optimize: Firstly,the swarm intelligence algorithm is used to optimize the threshold in the multi-threshold segmentation stage;Secondly,in the FCM segmentation stage,swarm intelligence algorithm is used to optimize the cluster center.Experiments show that the multi-strategy segmentation method based on CGNPSO is suitable for navel orange leaf disease image segmentation.(4)The feature extraction and optimization method of navel orange leaf disease image were proposedAiming at the disease atlas of navel orange leaves with small samples,a multi-angle feature extraction scheme based on color,texture and shape was proposed.Firstly,combined color spaces,RGB and HSV,to obtain comprehensive color information.Then,GLCM texture feature extraction method based on RI-LBP was used to extract texture features of disease images.RI-LBP has gray invariance,illumination invariance and rotation invariance.Extracting GLCM’s irrelevant features on the basis of RI-LBP image can not only provide more comprehensive texture information,but also solve the problem that RI-LBP histogram describes texture features with a large dimension.Finally,Hu moment and improved Zernike moment are used to extract the shape features of the disease image.Both Hu moment and improved Zernike moment have the invariance of rotation,translation and scaling,which can capture both global and local features of disease images.The above scheme comprehensively considers the robustness of the combined features,so as to effectively improve the accuracy of classification in the future.To solve the problem that the extracted combined features may have redundancy,a feature selection algorithm based on binary CGNPSO is proposed to optimize the features.The algorithm uses binary particle coding to represent the feature combination,and takes the minimization of classification error rate and the minimization of the number of features as the evaluation criteria of the feature subset,and adopts two-stage optimization according to the precocial decision strategy.Experimental results show that the algorithm can effectively reduce the redundant features in navel orange leaf disease images.(5)A disease recognition model of navel orange leaf based on swarm intelligence optimization was proposedOn the basis of navel orange disease image enhancement,image segmentation,feature extraction and optimization,three disease image classification models based on swarm intelligence optimization were proposed for small sample navel orange leaf disease atlas: ICS-SVM,CGNPSO-SVM and CGNPSO-NIN for disease image recognition.In ICS-SVM and CGNPSO-SVM models,ICS and CGNPSO were used to optimize SVM parameters to improve the classification accuracy.The CGNPSO-NIN model first uses gradient descent method to train the classical lightweight deep learning model NIN,and then takes the training results as the reference particle to establish the initial population,and use the CGNPSO algorithm to obtain the best NIN classifier.The experiments show that the CGNPSO-SVM model is suitable for the disease image recognition of navel orange leaves in small samples. |
