Project Proposal - Xception based Interpretable Architecture.pptx

An artificial Intelligence proposal for a Deep Interpretable Architecture for Plant Disease Detection

1. 1 RESEARCH PROPOSAL EMMANUEL OBOTU OBUTE STUDENT ID: 201123027   Department of Computer Science Faculty of Natural and Applied Science   Nile University of Nigeria Abuja, Nigeria   Under the Supervision of Professor Nwojo Agwu Dr. Ibrahim Salihu

2. XCEPTION BASED DEEP INTERPRETABLE ARCHITECTURE FOR PLANT DISEASE CLASSIFICATION

3. Introduction The agricultural sector remains the main source of food, raw materials and in some cases source of fuel, these in no small measure contributes to economic growth. As human population rises rapidly across the globe, so are the challenges facing agricultural industry, making it struggle to meet its obligation to our nations. Various conditions, including unfavorable climate change, crop diseases, lack of irrigation, industrialization, etc., endanger food security. Plant diseases pose the highest challenge to both quantitative and qualitative production of agricultural products. Plant diseases cause great damages to crops by significantly decreasing production. To ensure the quality and quantity of plant yields are guaranteed, it is important to take measures that protect plants from diseases or prevent the fast spread of such diseases. In order to ensure food security and the thriving of agricultural ecosystems, timely and accurate identification of plant diseases is crucial.  One of the active research fields in agriculture is the early diagnosis of plant diseases and its importance cannot be over-emphasized.   3

4. Introduction Cont.. The capabilities of Machine Learning and Computer Vision are being utilized in almost every discipline as they can give more promising results at lower costs. The agriculture industry is beginning to rely more on Deep Learning-based techniques for enhancement in this area. After revolutionizing, the domain of Computer Vision, Deep Learning can now solve many tasks such as automated plant disease diagnosis, soil fertility management, rainfall prediction, crop yield prediction, pest control, etc. Several Deep Learning-based techniques have been designed and implemented for the automatic detection of plant disease to help farmers increase plant productivity. The use of Convolutional Neural Networks (CNNs) to detect and classify diseases has been shown in various research. Compared to simplistic Machine Learning methods focused on custom features with which, it can be difficult to detect regions of interest, (ROIs) and to design and implement efficient parameters, this recent trend in CNN has created more precise classifiers. For plant disease classification, no doubt, these technologies produce very high accuracy but high precision is not enough. Users must have clarity on how the diagnosis is done and what symptoms exist in a particular plant. This can be attained with the help of image processing algorithms that ensure only the affected regions of a plant leaf are clearly highlighted. 4

5. Project Statement The Decoder uses the output representations of the image from the Teacher to reconstruct the image. The Student which is also a classifier that learns from this reconstructed image and classifies the disease with finer accuracy . T he visualization algorithm is built directly into the Teacher/Student architecture . In Teacher/Student, standard VGG16 architecture was used to implement the Teacher and the Student components. VGG16 is a linear architecture with no residual connections [8] and its accuracy is lesser compared to the newer architectures. The Xception (Extreme Inception) architecture will be implemented in the Teacher and Student components for higher accuracy and clear visuals. This research focuses on extending a visualization and classification architecture (Teacher/Student) based on combined learning of two deep classifiers that was proposed by Brahimi et al. 2019. Teacher/Student is a decipherable deep learning architecture that concurrently classifies the plant disease and as well visualizes inherent symptoms [7]. Teacher/Student uses an autoencoder for preserving only the salient features in the image. The autoencoder (Teacher & Decoder) handles the deconstruction and reconstruction of the image so that the noise is removed and only the important features remain in the image. The Teacher is a classifier that plays the role of an encoder in the autoencoder sub-architecture. 5

6. Aim & Objectives OBJECTIVES The main objectives of the research are as follows; • To develop an improved deep learning system for plant disease classification and visualization with Teacher/Student architecture using Xception for interpretability of classification decisions. • To develop web application with python Django framework that will serve as a means for taking input (images of selected plants) , classifying them as well as visualize features that supports the classification. • To improve the classification accuracy and visualization of plant disease inherent symptoms. AIM The aim of this research is to explore and implement an Xception based deep interpretable architecture for plant disease classification and visualization . 6

7. Literature Review 7 GAP Dubey et al. [11 ], Li et al. [9 ], Guan et al. [12] SVM, K-Means clustering, Bayesian and other orthodox image processing techniques have achieved certain results, with considerably high accuracy of disease recognition, but there are still deficiencies Recommendation it's pertinent to look out for more intelligent, rapid, and accurate disease recognition systems 1)The processes are cumbersome, time-consuming, highly subjective and tedious; 2) it heavily relies on artificial feature extraction; 3 ) it's difficult to check the disease recognition performance of the model or algorithm in additional complex environments

8. Literature Review Cont.. 8 GAP Park et al. [13 ], Montavon et al. [15 ] Convolutional Neural Networks, Region-based Convolutional Neural Network, Deep Neural Networks were used to classify plant diseases. Recommendation Introduction of algorithm for visualization of the representation learned by the classifiers. And use of architectures that supports multi-plant disease diagnosis These techniques function admirably, but one notable drawback of using these methods is that they lack transparency, which limits the interpretability of the outcomes.

9. Literature Review Cont.. 9 GAP Brahimi et al. 2019 Proposed the Teacher/Student architecture implemented the Teacher and Student components using Standard VGG-16 architecture. Recommendation Implementation of the Teacher and Student components using other architecture that uses residual/skip connections in encoder and decoder components VGG-16 is a linear architecture with no residual connections and its accuracy is lesser compared to the newer architectures .

10. Proposed Method We put forward a residual Teacher/Student architecture for classification as well as visualization, which is an improvement of the work done by Brahimi et al. [17] on Teacher/Student architecture for plant disease classification. To achieve this, we will implement the standard Xception (Extreme Inception) architecture as Teacher and Student. Residual or skip connections will be introduced along with batch normalization in the decoder section of the autoencoder which will help in achieving better accuracy for classification and visualization tasks. 10

11. Proposed Method According to the experiments performed by Kabir et al. [23], Xception architecture gives state-of-the-art precision when used for multi-plant disease diagnosis. This is because inter-layer outputs are modified by batch normalization into a standard format. For a particular data batch, batch normalization re-calibrates each of the data values depending on the mean and variance. The normalization of batches improves DNN architecture reliability and also contributes to faster convergence. Also, the residual connections sanction the gradients to flow through the architecture without going through non-linear activations. This in turn resolves the problem of vanishing or exploding gradients. 11

12. Xception vs other Classification Architectures Top-1 Accuracy Top-5 Accuracy Position VGG-16 0.715 0.901 3 Inception V3 0.782 0.941 2 Xception 0.790 0.945 1 12

13. Material T he segmented version of the PlantVillage dataset which is an open-source dataset, that contains 54,306 images categorized into 38 different classes of infected and healthy crops will be utilized, these classes constitute 14 plant species Python, Django framework Tensorflow framework Keras 13

14. 14 Any questions ?