There are a lot of proceduresthat employ computer vision for the detecting the disease in plants. In 9authors have presented an approach to detect disease using image processingtechniques for segmentation of the diseased spot. For the process of thedisease spot detection they have compared CIELB, YcbCr and HSI colour modelsand used Median filter for image smoothing. This provided an algorithm that successfullydetected diseases and remained independent of background noise.
Patil and Kumar in 10 describesa method for extracting colour and texture features of diseased leaf. Thetexture features such as correlation, energy, inertia and homogeneity are extractedby calculating the gray level co-occurrence matrix of an image. Also colourfeatures were extracted by obtaining the HSV of an image, together texture andcolour features were used to detect disease in maize leaves. The purpose of the feature selectionis to reduce the maximum number of irrelevant features while maintaining anacceptable classification accuracy 5. The computational efficiencysuperiority of PSO over the GA is statically proven. It appears that PSOoutperforms the GA with a larger differential in computational efficiency whenused to solve unconstrained nonlinear problems with continuous design variablesand less efficiency differential when applied to constrained nonlinear problemswith continuous or discrete design variables 7. In 11 the authors extractedcolour, edge and texture features from the diseased leaf image. Particle SwarmOptimization (PSO) was used for selection of features which would be used totrain deep forward neural network.
The proposed system was able to identifycotton diseases with an accuracy of 95%. In the past few years computervision and object recognition in particular has made substantial improvements.Large Scale Visual Recognition Challenge (ILSVRC) 16 which is based onImageNet database 17 is considered as benchmark for numerous computer visionrelated problems. AlexNet 2 a large, deep convolutional neural network forclassification of 1.2 million images into 1000 possible categories in ILSVRC,achieved a top-1 and top-5 error rates of 37.
5% and 17% which was better thanprevious state of the art. The authors in 4 employed deep learning forclassification of plant leaves diseases. They have used 30880 images forclassification of fifteen different types of classes. The images were used totrain a CaffeNet architecture which contains eight learning layers, fiveconvolutional and three fully connected layers. The network was able toclassify with an accuracy of 96.3%.
In another work the authors 3used different deep learning architectures for plant leaf diseaseclassification. The authors compared AlexNet and GoogLeNet architecture bytraining them from scratch as well as using transfer learning approach. Thenetwork was trained with PlantVillage dataset 12 which consisted of 54306 imagescontaining 38 classes of 14 crop species and 26 diseases. The overall accuracyvaried from 85.53% to 99.34% in this case GoogLeNet architecture was trainedusing transfer learning approach. Training large neural networks can be verytime-consuming process as it involves huge dataset. In our study we aim toemploy deep learning for the purpose of feature extraction.
All the abovementioned approaches used end to end neural networks, including for the purposeof classification. We present a computationally optimal approach which can beeasily used on smartphones. We employ deep neural network for featureextraction and Particle Swarm Optimization for the feature selection. Theoptimal set of features is used to train a classifier for the classifying imagesinto their respective classes.