REQUIREMENT COLLECTION FORM
Note: Read the Guidelines carefully, before submitting.
Name Fill by student
Contact Fill by student
College Name Fill by student
Project Title Image Segmentation Based Fish Wound and Lesion Detection System
Using U-Net for Aquaculture Disease Analysis
Base Paper Link
Expected Inputs A single high-resolution RGB fish body image (JPG or PNG format)
uploaded by the user through a web interface. The image should clearly
show the fish skin surface. Example: A photo of a freshwater fish
showing red spots or white tail discoloration on its body.
Expected Outputs Binary segmentation mask showing the fish body separated from
background
Disease class label (example: Bacterial Aeromoniasis, Fungal
Saprolegniasis, White Tail Disease, Healthy etc.)
Lesion area percentage (example: 23% of fish body is infected)
Color overlay visualization (GradCAM) — healthy region in green,
infected/lesion region in red
Dataset (If Two separate Kaggle datasets are used. They are NOT merged. Each
Available) dataset trains a different model independently.
Dataset 1 — Large Scale Fish Dataset (Kaggle)
Link: [Link]
dataset
Purpose: Trains the U-Net model to segment fish body from
background. This dataset already has 9000 fish images with paired
pixel-level ground truth body masks ready to use.
Dataset 2 — Freshwater Fish Disease Aquaculture in South Asia
(Kaggle)
Link: [Link]
disease-aquaculture-in-south-asia
Purpose: Trains the CNN classifier to identify which disease is present.
Contains 1750 images across 7 disease classes.
Algorithm (If Attention U-Net (Enhanced U-Net with Attention Gates)
Known) Base architecture is U-Net as mentioned in the project title. Enhanced
with Attention Gates on skip connections so the model focuses only on
the lesion/disease region and ignores background noise. Encoder uses
pretrained ResNet-34. Loss function is Dice Loss combined with Binary
Cross Entropy. A separate CNN classifier runs after U-Net to identify the
disease class.
Tech Stack: Frontend: HTML, CSS, JavaScript — user uploads fish image and sees
results
(In case of Full Backend: Flask (Python) — handles image upload, runs both models,
stack project) returns mask and disease class
Model: PyTorch with segmentation-models-pytorch library
Communication: REST API between frontend and Flask backend
Deadline To be decided
Comments Dataset can be change if something is not feasible to do.
Workflow User uploads a single fish image through the web interface
Flask backend receives the image and passes it to Model 1
(Attention U-Net trained on Large Scale Fish Dataset) which
segments and isolates the fish body from the background using
pixel-level mask
The isolated fish body region is then passed to Model 2 (CNN
Classifier trained on Freshwater Fish Disease dataset) which
predicts the disease class and confidence score
GradCAM is applied on Model 2 output to generate a heatmap
highlighting the disease-concentrated region. This heatmap is
overlaid on the fish body mask to produce the green/red
visualization and calculate lesion area percentage
Flask sends back the mask image, disease label, lesion
percentage and overlay to the frontend
Frontend displays all results to the user on a clean result page
Points to Note:
● The project development strictly adheres to the provided algorithm and workflow
● Changes to the above-mentioned entities results in the entire reconstruction of the
project
● Leave the cells blank if you do not have the necessary data
● Explain everything in detail, requirements that are not mentioned in the form will cost
extra as per requirement.
● Dataset and algorithms will not be changed later once a project development is started