ABOUT MTECH PROJECT
1. "Why are you using two models instead of just one?"
The Simple Answer: "Most systems today have a trade-off: they are either fast enough for a
webcam but not very accurate, or they are very accurate but too slow to run in real-time. My
project is a hybrid approach. I’m developing one lightweight model for instant webcam
feedback and a second, more powerful transformer model for high-accuracy video file
analysis."
2. "What exactly is MediaPipe doing in your project?"
The Simple Answer: "Think of MediaPipe as a tool that 'simplifies' the human body. Instead
of the computer trying to look at every pixel in a video frame, MediaPipe just picks out the
important 'dots' or landmarks on the face, hands, and body. This makes the data much smaller
and easier for my model to process quickly."
3. "Why did you add an 'Attention Mechanism' to the BiLSTM?"
The Simple Answer: "A BiLSTM looks at the movement over time. The 'Attention
Mechanism' acts like a spotlight. It helps the model ignore the background and focus only on
the most important frames or movements—like a specific hand gesture—that define the
activity."
4. "What is VideoMAE, and why not just use a normal CNN?"
The Simple Answer: "VideoMAE is a modern 'Transformer' model specifically built for
video. Standard CNNs often struggle to understand how an action changes over time.
VideoMAE is much better at understanding the 'flow' of a video. Since it’s a heavy model,
I’m using transfer learning to fine-tune it so I don't have to train it from scratch."
5. "What is the 'extra' or unique thing you are doing here?"
The Simple Answer: "Two things. First, most projects only look at body movement, but I am
including facial expressions in the same system. Second, I am integrating both real-time and
file-based recognition into one Flask web app. This gives the user a complete tool instead of
just a single script."
6. "Since you are still developing this, what is your biggest focus right now?"
The Simple Answer: "Right now, I am focusing on the integration phase. I want to make sure
the Flask application can handle the switch between the webcam and the video upload
smoothly without crashing or slowing down the computer."
Be honest about the status: If they ask for a specific accuracy percentage, say: "I am currently
in the training and evaluation phase, so I am still fine-tuning the models to reach the best
possible accuracy".
The "Why": If they ask why you chose this, just repeat: "I wanted to solve the balance
between speed and accuracy".
ABOUT MINI PROJECT
My project is titled 'Mulberry Leaf Disease Detection Using CNN-Based Smart Android
Application'.
The motivation behind this work comes from the sericulture industry, where mulberry leaves
are the sole food source for silkworms. Diseases like Leaf Spot and Leaf Rust significantly
degrade the quality of these leaves, which directly impacts silk production. My goal was to
move away from slow, manual inspection and create an automated tool using Deep Learning
to identify these diseases accurately."
2. Dataset Collection & Preparation (3 Minutes)
"For this project, I used a custom-labeled image dataset.
Data Collection: I gathered images representing three specific classes: Healthy
leaves, Leaf Spot, and Leaf Rust.
Preprocessing: To prepare the data for the model, I resized all images to a standard
$124 \times 124$ pixels.
Data Augmentation: To ensure the model doesn't overfit and can handle different
image variations, I applied techniques like rotation, zooming, and flipping.
Data Splitting: I organized the data into a 60-20-20 split—60% for training, 20% for
validation during the training process, and 20% for final testing to verify accuracy."
3. The Technical Approach (3 Minutes)
"I utilized Convolutional Neural Networks (CNNs) because they are highly effective at
extracting spatial features from images.
Transfer Learning: Instead of training a model from scratch, I used Transfer
Learning with pre-trained architectures like MobileNetV3-Small, VGG19, and
ResNet50.
Why MobileNet? I prioritized MobileNet because it is a lightweight architecture
specifically optimized for mobile devices, which was crucial for my Android
application.
Classification: I added custom Dense Layers (with 512 and 256 neurons) followed
by a SoftMax output layer to classify the images into the three categories mentioned
earlier."
4. Explainability & Implementation (2 Minutes)
"A unique feature of this project is the use of Grad-CAM Visualization.
Grad-CAM: This allows us to see 'why' the model made a certain decision by
highlighting the specific areas of the leaf that the model focused on. It makes the AI
more transparent and reliable.
Deployment: Finally, I integrated the best-performing model into a Smart Android
Application. This app allows a user to upload an image from the dataset and receive
an immediate classification of the leaf's health.
Outcome: The system provides a lightweight, accurate solution that helps in early
disease detection, ultimately supporting better crop management in sericulture."
Important Interview "Save" for the Dataset Question:
If they ask, "Where did you get the dataset?"
"I used a custom dataset specifically prepared for this study. I focused on ensuring the
images were correctly labeled into the three categories (Healthy, Spot, Rust) so that the CNN
could learn the distinct visual patterns of each fungal infection."
1. "Why did you choose MobileNetV3 for this project instead of a larger model like
VGG19 or ResNet50?"
Answer:
The Technical Reason: While VGG19 and ResNet50 are highly accurate, they have
millions of parameters and a very large file size, which makes them slow on mobile
devices.
The Practical Choice: I chose MobileNetV3-Small because it is specifically
designed to be lightweight and efficient for mobile hardware.
The Outcome: It offers a better trade-off between speed and accuracy, ensuring the
Android app runs smoothly without needing high-end hardware.
2. "You mentioned you used a 'Custom Dataset.' How did you ensure the quality and
reliability of this data?"
Answer:
Labeling: I ensured that the images were accurately categorized into three distinct
classes: Healthy, Leaf Spot, and Leaf Rust.
Diversity: By using Data Augmentation (rotation, flipping, and zooming), I
expanded the dataset to help the model recognize leaves in different positions and
lighting conditions.
Preprocessing: Every image was standardized to $124 \times 124$ pixels to ensure
consistent input for the CNN.
3. "Explain the concept of 'Transfer Learning' as if you were teaching a first-year
student."
Answer:
The Analogy: Imagine someone who already knows how to drive a car; they will
learn to drive a bus much faster than someone who has never driven at all.
The Project Application: Similarly, I used models like ResNet50 that were already
pre-trained on the ImageNet dataset (millions of general images).
The Benefit: The model already knows how to identify basic shapes and textures. I
simply "fine-tuned" it to recognize the specific patterns of mulberry leaf diseases.
4. "What is the role of Grad-CAM in your application, and why is it necessary?"
Answer:
The Goal: CNNs are often considered "black boxes." Grad-CAM (Gradient-
weighted Class Activation Mapping) is used to make the model's decision-making
transparent.
The Visual Proof: It generates a heatmap that highlights the exact area of the leaf the
model focused on to make its prediction.
User Trust: This helps a farmer trust the app because they can see that the AI is
actually looking at the disease spots and not just a random part of the leaf.
5. "What are the limitations of this project, and how would you improve it in the
future?"
Answer:
Current Limitation: The project currently relies on a fixed dataset of specific
mulberry diseases (Leaf Spot and Leaf Rust).
Future Improvement: I would like to expand the dataset to include more types of
pests and nutritional deficiencies.
Real-world Deployment: I would also focus on making the model even more robust
to handle very low-light conditions or extremely blurry photos taken in the field.