Week 2— Plant Disease Detection
Theme: Classification to plants that healthy or diseased and predict to photographed plant disease.
Team Members: Fatmanur Turhan, Sevda Sayan, İsmet Seyhan
Convolutional Neural Network (CNN)
The convolutional neural network (CNN or ConvNet) is a deep, feedforward artificial neural network commonly used for the analysis of visual images. CNN uses a variety of multi-layer detectors designed to require minimal pretreatment. They use relatively little preprocessing compared to other image classification algorithms.
CNN has many architectures and has different layer numbers and different activation functions.
Convolutional Layer: Convolution is a multiplication of a large image with a smaller image. This multiplication and addition are performed for each step by scrolling on the large image.
Pooling Layer: A filter of a certain size is run around on the large image. The process is to select a single number of the big picture where the filter travels. How this number is selected determines the type of filter. For example:
If the maximum number is selected, “max pooling”, if the numbers are averaged, the average pooling layer is used.
Flattening Layer: Prepares data for the neural network by reducing the input to a one-dimensional state.
Fully-Connected Layer: It consists of neurons that are fully connected to the previous layer. It describes a classical artificial neural network.
Design and Architecture
According to the results of the articles mentioned in our previous blog (please click Week1), 54,306 images of plant leaves, which have a spread of 38 class labels assigned to them are analyzed. Each class label is a crop-disease pair and makes an attempt to predict the crop-disease pair given just the image of the plant leaf. In all the approaches described in this paper, the images to 256 × 256 pixels are resized and both the model optimization and predictions on these downscaled images are performed.
Three different versions are used for the whole PlantVillage dataset. Firstly, with the PlantVillage dataset as it is, in color; secondly, experiments with a gray-scaled version of the PlantVillage dataset, and finally, experiments where the leaves were segmented. (Hence removing all the extra background information which might have the potential to introduce some inherent bias in the dataset due to the regularized process of data collection in case of PlantVillage dataset)
This set of experiments was designed to understand if the neural network actually learns the “notion” of plant diseases, or if it is just learning the inherent biases in the dataset.
According to our reference paper, we have observed that the best results in the classification of plant diseases are obtained by AlexNet and GoogleNet architectures. So we decided to develop our project using AlexNet or GoogleNet. We will try these two architectures and which architecture gives the best results, then we choose that architecture. Now let’s look at the structures of these two architectures.
AlexNet Architecture
ReLU (Rectified Linear Unit) is used as an activation function in AlexNet architecture and max-pooling is used in pooling layers. Large and deep, this network model is a two-part model on a parallel dual GPU. Approximately 60 million parameters are calculated. In the ImageNet ILSVRC competition, the classification accuracy is a breakpoint in the image classification problem, providing a sudden increase from 74.3% to 83.6%.
GoogleNet Architecture
GoogleNet architecture consists of modules. Each module consists of different dimensional convolution and max-pooling processes. Thus, with a more complex network model, less calculation and faster design are done. In the GoogLeNet model, the model itself is expanding.
Thank you for reading…
References:
