Article Preview
Top2. Literature Survey
To understand what different methods were used to identifying diseases in different plants, a summary of the different papers need to be understood. On the Plant Village dataset, the authors (Tm, P., Pranathi et al., 2018) developed a method to categorize sick tomato leaves into ten distinct groups. Data gathering, data pre-processing, and classification are the three phases in their suggested technique. Because the dataset contained low-noise pictures, noise reduction was not necessary for the pre-processing stage, and the images were scaled to 60 × 60 resolution. The Z-score technique, which is the mean and standard deviation, was used to normalize the pictures. They utilized Convolutional Neural Networks (CNNs) using several deep learning architectures such as LeNet, AlexNet, and GoogleNet for classification. The LeNet architecture produced the greatest results, with an accuracy of 94.8 percent. However, because the pictures do not have any noisy backgrounds, this approach is computationally intensive, and its practical practicality is debatable.
In 2019 (Wang, Q. et al., 2019) researchers merged three deep convolutional neural networks, VGG169, ResNet50, and ResNet101, which are commonly used for extracting features, with a Faster RCNN structure to diagnose tomato illnesses. The entire procedure is broken down into four steps: The Deep Convolutional Neural Networks (DCNN) is the first, and it is used to extract feature maps from input pictures. The (Region Proposal Networks) RPN is generated in the second phase. The ROI (Region of interest) pooling stage involves RPN obtaining a proposed feature map of fixed size and DCNN obtaining the final feature map, which is then fed into the complete connection layer at the back for target detection and placement. The feature map is then fed into the complete connection layer, with the SoftMax layer utilized to precisely categorize the input. Simultaneously, the boundary box regression procedure is performed to determine the exact position of the tomato fruit. resnet101 is proven to have the best overall performance. However, due to the limited sample size utilized for training, the model in the article has a low detection accuracy. As a result, work on upgrading the network structure is required.