import streamlit as st
from ultralytics import YOLO
import tensorflow as tf # Change this to import TensorFlow
import numpy as np
from PIL import Image, ImageOps, ImageDraw, ImageFont
import pandas as pd
import time
from paddleocr import PaddleOCR, draw_ocr
import re
import dateparser
import os
import matplotlib.pyplot as plt
#######Llama3bi integration########
import torch
from transformers import pipeline
model_id = "meta-llama/Llama-3.2-3B-Instruct"
pipe = pipeline(
"text-generation",
model=model_id,
torch_dtype=torch.bfloat16,
device_map="auto",
)
messages = [
{"role": "system", "content": """Your task is to get the product details out of the text given. The text given will be raw text from OCR of social media images of products,
and the goal is to get product details and description so that it can be used for amazon product listing. """},
]
# Function to get Instagram post details
import instaloader
def get_instagram_post_details(post_url):
try:
shortcode = post_url.split('/')[-2]
post = instaloader.Post.from_shortcode(L.context, shortcode)
# Retrieve caption and image URL
caption = post.caption
image_url = post.url
return caption, image_url
except Exception as e:
return str(e), None
# Initialize PaddleOCR model
ocr = PaddleOCR(use_angle_cls=True, lang='en')
# Team details
team_members = [
{"name": "Aman Deep", "image": "aman.jpg"}, # Replace with actual paths to images
{"name": "Nandini", "image": "myimage.jpg"},
{"name": "Abhay Sharma", "image": "gaurav.jpg"},
{"name": "Ratan Prakash Mishra", "image": "anandimg.jpg"}
]
# Function to preprocess the images for the model
def preprocess_image(image):
"""
Preprocess the input image for model prediction.
Args:
image (PIL.Image): Input image in PIL format.
Returns:
np.ndarray: Preprocessed image array ready for prediction.
"""
try:
# Resize image to match model input size
img = image.resize((128, 128), Image.LANCZOS) # Using LANCZOS filter for high-quality resizing
# Convert image to NumPy array
img_array = np.array(img)
# Check if the image is grayscale and convert to RGB if needed
if img_array.ndim == 2: # Grayscale image
img_array = np.stack([img_array] * 3, axis=-1) # Convert to 3-channel RGB
elif img_array.shape[2] == 1: # Single-channel image
img_array = np.concatenate([img_array, img_array, img_array], axis=-1) # Convert to RGB
# Normalize pixel values to [0, 1] range
img_array = img_array / 255.0
# Add batch dimension
img_array = np.expand_dims(img_array, axis=0) # Shape: (1, 128, 128, 3)
return img_array
except Exception as e:
print(f"Error processing image: {e}")
return None # Return None if there's an error
# Function to create a high-quality circular mask for an image
def make_image_circular1(img, size=(256, 256)):
img = img.resize(size, Image.LANCZOS)
mask = Image.new("L", size, 0)
draw = ImageDraw.Draw(mask)
draw.ellipse((0, 0) + size, fill=255)
output = ImageOps.fit(img, mask.size, centering=(0.5, 0.5))
output.putalpha(mask) # Apply the mask as transparency
return output
# Function to check if a file exists
def file_exists(file_path):
return os.path.isfile(file_path)
def make_image_circular(image):
# Create a circular mask
mask = Image.new("L", image.size, 0)
draw = ImageDraw.Draw(mask)
draw.ellipse((0, 0, image.size[0], image.size[1]), fill=255)
# Apply the mask to the image
circular_image = Image.new("RGB", image.size)
circular_image.paste(image.convert("RGBA"), (0, 0), mask)
return circular_image
# Function to extract dates from recognized text using regex
def extract_dates_with_dateparser(texts, result):
date_texts = []
date_boxes = []
date_scores = []
def is_potential_date(text):
valid_date_pattern = r'^(0[1-9]|[12][0-9]|3[01])[-/.]?(0[1-9]|1[0-2])[-/.]?(\d{2}|\d{4})$|' \
r'^(0[1-9]|[12][0-9]|3[01])[-/.]?[A-Za-z]{3}[-/.]?(\d{2}|\d{4})$|' \
r'^(0[1-9]|1[0-2])[-/.]?(\d{2}|\d{4})$|' \
r'^[A-Za-z]{3}[-/.]?(\d{2}|\d{4})$'
return bool(re.match(valid_date_pattern, text))
dates_found = []
for i, text in enumerate(texts):
if is_potential_date(text): # Only process texts that are potential dates
parsed_date = dateparser.parse(text, settings={'DATE_ORDER': 'DMY'})
if parsed_date:
dates_found.append(parsed_date.strftime('%Y-%m-%d')) # Store as 'YYYY-MM-DD'
date_texts.append(text) # Store the original text
date_boxes.append(result[0][i][0]) # Store the bounding box
date_scores.append(result[0][i][1][1]) # Store confidence score
return dates_found, date_texts, date_boxes, date_scores
# Function to display circular images in a matrix format
def display_images_in_grid(images, max_images_per_row=4):
num_images = len(images)
num_rows = (num_images + max_images_per_row - 1) // max_images_per_row # Calculate number of rows
for i in range(num_rows):
cols = st.columns(min(max_images_per_row, num_images - i * max_images_per_row))
for j, img in enumerate(images[i * max_images_per_row:(i + 1) * max_images_per_row]):
with cols[j]:
st.image(img, use_column_width=True)
# Function to display team members in circular format
def display_team_members(members, max_members_per_row=4):
num_members = len(members)
num_rows = (num_members + max_members_per_row - 1) // max_members_per_row # Calculate number of rows
for i in range(num_rows):
cols = st.columns(min(max_members_per_row, num_members - i * max_members_per_row))
for j, member in enumerate(members[i * max_members_per_row:(i + 1) * max_members_per_row]):
with cols[j]:
img = Image.open(member["image"]) # Load the image
circular_img = make_image_circular(img) # Convert to circular format
st.image(circular_img, use_column_width=True) # Display the circular image
st.write(member["name"]) # Display the name below the image
# Title and description
st.title("Amazon Smbhav")
# Team Details with links
st.sidebar.title("Amazon Smbhav")
st.sidebar.write("DELHI TECHNOLOGICAL UNIVERSITY")
# Navbar with task tabs
st.sidebar.title("Navigation")
st.sidebar.write("Team Name: sadhya")
app_mode = st.sidebar.selectbox("Choose the task", ["Welcome","Project Details", "Task 1","Team Details"])
if app_mode == "Welcome":
# Navigation Menu
st.write("# Welcome to Amazon Smbhav! 🎉")
# Example for adding a local video
video_file = open('Finalist.mp4', 'rb') # Replace with the path to your video file
video_bytes = video_file.read()
# Embed the video using st.video()
st.video(video_bytes)
# Add a welcome image
welcome_image = Image.open("grid_banner.jpg") # Replace with the path to your welcome image
st.image(welcome_image, use_column_width=True) # Display the welcome image
elif app_mode=="Project Details":
st.markdown("""
## Navigation
- [Project Overview](#project-overview)
- [Proposal Round](#proposal-round)
- [Problem Statement](#problem-statement)
- [Proposed Solution](#proposed-solution)
""")
# Project Overview
st.write("## Project Overview:")
st.write("""
### Problem Statement
_Develop a system that automates Amazon product listings from social media content, extracting and organizing details from posts to generate accurate, engaging, and optimized listings._
---
### Solution Overview
Our system simplifies the listing process by analyzing social media content, using OCR, image recognition, LLMs, and internet data to create professional Amazon listings.
---
### Task Breakdown
#### Task 1: OCR for Image and Label Details
**Objective:** Extract core product details from images, labels, and packaging found in social media posts.
- **Tools:** PaddleOCR, LLMs.
- **Approach:**
- Use PaddleOCR to scan images for text, identifying product names, brands, and key features.
- Apply LLMs to refine extracted data, categorize key information (product name, type, features), and enhance product descriptions.
- Integrate internet sources to cross-verify product details, retrieve additional information, and collect metadata like the brand background or product specs.
---
#### Additional Task: Image Recognition & Object Counting
**Objective:** Quantify objects within social media images for batch products or multi-item listings.
- **Tools:** YOLOv8.
- **Approach:**
- Train YOLOv8 on a relevant dataset to recognize specific product types or packaging layouts.
- Use object detection counts to provide quantitative data (e.g., "3-item bundle"), enhancing accuracy in listings.
---
#### Task 2: Data Validation & Structuring
**Objective:** Organize and validate extracted information, ensuring it’s formatted to meet Amazon’s listing requirements.
- **Tools:** Regex, LLMs.
- **Approach:**
- Format and validate extracted details into Amazon-compliant structures (titles, descriptions, bullet points).
- Use regex and parser tools for accuracy checks.
- Leverage LLMs to create compelling descriptions and marketing brochures.
- Search online for supplementary media (images/videos) to enrich the listing.
---
#### Task 3: Amazon API Integration
**Objective:** Connect with Amazon’s API to publish fully formed product listings directly.
- **Tools:** Amazon MWS or Selling Partner API.
- **Approach:**
- Send structured listing data (text, media, product details) to Amazon’s API endpoints.
- Handle feedback for submission errors and make necessary adjustments.
- Develop a UI/dashboard for users to preview and edit listings before publishing.
---
### Future Enhancements
- **Model Improvement:** Further refine OCR and parsing accuracy.
- **Dashboard Development:** Enable users to preview and customize listings.
- **Multi-Market Compatibility:** Expand support to other e-commerce platforms.
This approach automates listing creation directly from social media content, helping sellers quickly launch optimized Amazon product pages.
""")
elif app_mode == "Team Details":
st.write("## Meet Our Team:")
display_team_members(team_members)
st.write("Delhi Technological University")
elif app_mode == "Task 1":
st.write("## Task 1: 🖼️ OCR to Extract Details 📄")
st.write("Using OCR to extract details from product packaging material, including brand name and pack size.")
# Instantiate Instaloader
L = instaloader.Instaloader()
# Streamlit UI
st.title("Instagram Post Details Extractor")
# Text input for Instagram post URL
post_url = st.text_input("Enter Instagram Post URL:")
if post_url:
caption, image_path = get_instagram_post_details(post_url)
if image_path and os.path.exists(image_path):
st.subheader("Caption:")
st.write(caption)
st.subheader("Image:")
# Load and display the image
image = Image.open(image_path)
st.image(image, use_column_width=True)
# Convert image to numpy array for OCR processing
img_array = np.array(image)
# Perform OCR on the image
st.write(f"Extracting details from {uploaded_image.name}...")
result = ocr.ocr(img_array, cls=True)
#############################
#OCR result text to be parsed here through LLM and get product listing content.
messages.append({"role": "user", "content": ""})
outputs = pipe(
messages,
max_new_tokens=256,
)
productListingContent = outputs[0]["generated_text"][-1]
st.markdown(productListingContent)
#############################
# Clean up (optional)
os.remove(image_path)
else:
st.error("Failed to retrieve the post details. Please check the URL.")
# File uploader for images (supports multiple files)
uploaded_files = st.file_uploader("Upload images of products", type=["jpeg", "png", "jpg"], accept_multiple_files=True)
if uploaded_files:
st.write("### Uploaded Images in Circular Format:")
circular_images = []
for uploaded_file in uploaded_files:
img = Image.open(uploaded_file)
circular_img = make_image_circular(img) # Create circular images
circular_images.append(circular_img)
# Display the circular images in a matrix/grid format
display_images_in_grid(circular_images, max_images_per_row=4)
# Function to simulate loading process with a progress bar
def simulate_progress():
progress_bar = st.progress(0)
for percent_complete in range(100):
time.sleep(0.02)
progress_bar.progress(percent_complete + 1)
# Function to remove gibberish using regex (removes non-alphanumeric chars, filters out very short text)
def clean_text(text):
# Keep text with letters, digits, and spaces, and remove short/irrelevant text
return re.sub(r'[^a-zA-Z0-9\s]', '', text).strip()
# Function to extract the most prominent text (product name) and other details
def extract_product_info(results):
product_name = ""
product_details = ""
largest_text_size = 0
for line in results:
for box in line:
text, confidence = box[1][0], box[1][1]
text_size = box[0][2][1] - box[0][0][1] # Calculate height of the text box
# Clean the text to avoid gibberish
clean_text_line = clean_text(text)
if confidence > 0.7 and len(clean_text_line) > 2: # Only consider confident, meaningful text
if text_size > largest_text_size: # Assume the largest text is the product name
largest_text_size = text_size
product_name = clean_text_line
else:
product_details += clean_text_line + " "
return product_name, product_details.strip()
if st.button("Start Analysis"):
simulate_progress()
# Loop through each uploaded image and process them
for uploaded_image in uploaded_files:
# Load the uploaded image
image = Image.open(uploaded_image)
# st.image(image, caption=f'Uploaded Image: {uploaded_image.name}', use_column_width=True)
# Convert image to numpy array for OCR processing
img_array = np.array(image)
# Perform OCR on the image
st.write(f"Extracting details from {uploaded_image.name}...")
result = ocr.ocr(img_array, cls=True)
#############################
#OCR result text to be parsed here through LLM and get product listing content.
messages.append({"role": "user", "content": ""})
outputs = pipe(
messages,
max_new_tokens=256,
)
productListingContent = outputs[0]["generated_text"][-1]
st.markdown(productListingContent)
#############################
# Process the OCR result to extract product name and properties
product_name, product_details = extract_product_info(result)
# UI display for single image product details
st.markdown("---")
st.markdown(f"### **Product Name:** `{product_name}`")
st.write(f"**Product Properties:** {product_details}")
st.markdown("---")
else:
st.write("Please upload images to extract product details.")
def make_image_circular1(image):
# Create a circular mask
mask = Image.new("L", image.size, 0)
draw = ImageDraw.Draw(mask)
draw.ellipse((0, 0, image.size[0], image.size[1]), fill=255)
# Apply the mask to the image
circular_image = Image.new("RGB", image.size)
circular_image.paste(image.convert("RGBA"), (0, 0), mask)
return circular_image
def display_images_in_grid1(images, max_images_per_row=4):
rows = (len(images) + max_images_per_row - 1) // max_images_per_row # Calculate number of rows needed
for i in range(0, len(images), max_images_per_row):
cols_to_show = images[i:i + max_images_per_row]
# Prepare to display in a grid format
cols = st.columns(max_images_per_row) # Create columns dynamically
for idx, img in enumerate(cols_to_show):
img = img.convert("RGB") # Ensure the image is in RGB mode
if idx < len(cols):
cols[idx].image(img, use_column_width=True)
# Footer with animation
st.markdown("""
""", unsafe_allow_html=True)