Grocery Store Product Classification

This repository contains the implementation of a neural network that classifies smartphone pictures of grocery store products into one of 43 predefined categories. This project is divided into two parts:

1. Implementing a custom neural network from scratch.
2. Fine-tuning a pretrained ResNet-18 model using PyTorch.

Project Overview

Dataset

The dataset used for this project contains natural images of products taken with a smartphone camera in grocery stores. It is divided into three splits:

• Train
• Validation (Val)
• Test

The dataset includes images belonging to the following 43 product categories:

0.  Apple            1.  Avocado           2.  Banana          3.  Kiwi
4.  Lemon            5.  Lime              6.  Mango           7.  Melon
8.  Nectarine        9.  Orange           10. Papaya         11. Passion-Fruit
12. Peach           13. Pear             14. Pineapple      15. Plum
16. Pomegranate     17. Red-Grapefruit   18. Satsumas       19. Juice
20. Milk            21. Oatghurt         22. Oat-Milk       23. Sour-Cream
24. Sour-Milk       25. Soyghurt         26. Soy-Milk       27. Yoghurt
28. Asparagus       29. Aubergine        30. Cabbage        31. Carrots
32. Cucumber        33. Garlic           34. Ginger         35. Leek
36. Mushroom        37. Onion            38. Pepper         39. Potato
40. Red-Beet        41. Tomato           42. Zucchini

Goals

Part 1: Design a Custom Neural Network

• Implement a convolutional neural network (CNN) for image classification from scratch.
• Aim to achieve around 60% validation accuracy.
• Justify all design choices, including:
  • Network architecture
  • Training hyperparameters
  • Dataset preprocessing steps
• Document results and improvements using training plots, tables, or console outputs.

Part 2: Fine-Tune a Pretrained Network

• Fine-tune a ResNet-18 model (pretrained on ImageNet-1K) using PyTorch.
• Fine-tuning steps:
  1. Train the model with the same hyperparameters used in Part 1.
  2. Optimize training hyperparameters to achieve a validation accuracy between 80% and 90%.

Results

• Part 1: Achieved 62.16% accuracy with a custom CNN.
• Part 2: Achieved 88.51% accuracy by fine-tuning ResNet-18.