Raw Material Classification Using Deep Learning

Project Definition

In this project, we aim to achieve raw material classification using pre-trained AI image classification models with transfer learning.
The models are implemented using TensorFlow, and the selected architectures are:

• EfficientNet-B1
• MobileNetV2
• InceptionV3

The system classifies images into four material classes:

• Wood
• Glass
• Plastic
• Metal

The objective is to enable the model to learn texture and lighting patterns that generalize well to real-life images.

---

Inception Architecture (Main Model)

Introduction

InceptionV3, introduced by Google, follows a different design philosophy compared to traditional deep networks.
Instead of increasing depth, Inception focuses on increasing width by applying multiple filter sizes simultaneously.

Traditional models force a single kernel size per layer, while Inception applies:

• 1×1
• 3×3
• 5×5 convolutions
  in parallel.

---

Mechanism (Split–Transform–Merge)

1. Parallel Branching
   Multiple convolution filters (1×1, 3×3, 5×5) operate in parallel on the same input.

2. Concatenation
   Outputs from all branches are concatenated depth-wise.

This allows the network to capture fine details and broad spatial features simultaneously.

---

Main Concepts in InceptionV3

1. Dimensionality Reduction (1×1 Convolutions)

• 1×1 convolutions are used as bottleneck layers
• They reduce channel depth before expensive operations
• This prevents computational bottlenecks

2. Spatial Factorization (V3 Specialty)

• Replaces costly 5×5 convolutions with two stacked 3×3 convolutions
• Reduces parameters by ~28%
• Preserves receptive field while increasing non-linearity

InceptionV3 Visual Module

mindmap
  root((InceptionV3))
    Philosophy
      Wider not Deeper
      Multi-scale Feature Extraction
    Inception Module
      Parallel Branches
        1x1 Convolution
        3x3 Convolution
        5x5 Convolution
      Concatenation
        Feature Map Merging
    Key Concepts
      Dimensionality Reduction
        1x1 Convolutions
        Bottleneck Layers
      Spatial Factorization
        5x5 to Two 3x3
        Reduced Parameters
        Same Receptive Field

Loading

---

MobileNetV2 & EfficientNet-B1

A. Summaries

1. MobileNetV2

A lightweight architecture designed for mobile and edge devices.

Key Innovations:

• Inverted Residuals
  Structure: Narrow → Wide → Narrow
• Linear Bottlenecks
  Removes ReLU in narrow layers to preserve information

---

2. EfficientNet-B1

EfficientNet demonstrates that high accuracy does not require massive models.

Core Ideas:

• Compound Scaling (Φ)
  Uniform scaling of depth, width, and resolution
• MBConv Blocks
  Based on MobileNetV2 inverted residuals
• Squeeze-and-Excitation (SE)
  Allows the model to focus on important feature channels

---

B. Architecture Comparisons

InceptionV3 vs EfficientNet

• InceptionV3: Manually hand-crafted architecture
• EfficientNet: Discovered using Neural Architecture Search (NAS)
• EfficientNet is cleaner and easier to train

InceptionV3 vs MobileNetV2

• EfficientNet acts as a "Super MobileNet"
• Adds SE layers for higher accuracy with minimal extra cost

---

Model Comparison Table

Feature	MobileNetV2	InceptionV3	EfficientNet-B1
Release Year	2018	2015	2019
Primary Focus	Speed & Latency	High Accuracy	Accuracy vs Size Balance
Model Size	~3.4M Params	~24M Params	~7.8M Params
Core Block	Inverted Residual	Inception Module	MBConv + SE
Key Innovation	Linear Bottlenecks	Filter Factorization	Compound Scaling
Compute Cost	Very Low	High (GPU)	Low–Medium
Best Use Case	Mobile / Edge Devices	Server-side Accuracy	General-purpose

---

Dataset Aggregation & Curation

To improve generalization, a multi-source dataset was constructed.

Data Sources

• MINC-2500
  High-resolution texture-focused dataset
• Garbage Classification Datasets
  Real-world noise and deformations
• Google Images (Web Scraping)
  Used to balance classes and increase diversity

Data Cleaning

• Manual removal of:
  • Duplicates
  • Watermarks
  • Low-quality images

---

Challenges & Obstacles

1. Synthetic Overfitting

• Initial dataset contained 3D-rendered images
• Achieved misleading 99–100% accuracy
• Failed in real-world scenarios
• Dataset was discarded completely

2. Data Scarcity

• No unified high-quality raw material dataset
• Required extensive manual aggregation and validation

3. Model Selection

Evaluated multiple architectures:

• ResNet (18, 50)
• ConvNeXt
• MobileNet
• Inception
• EfficientNet (B0, B1)

Goal: balance accuracy vs efficiency

4. Optimization

• Initial training plateaued below target accuracy
• Hyperparameter tuning + architectural analysis
• Final models exceeded 80% accuracy

---

Results (Best Model: InceptionV3)

xychart-beta
    title "Training vs Validation vs Test Accuracy"
    x-axis ["Best Training","Best Validation","Final Test"]
    y-axis "Accuracy (%)" 0 --> 100
    bar [91.2, 88.1, 86.6]

Loading

---

Class-wise Accuracy (Training Phase)

xychart-beta
    title "Class-wise Accuracy (Training Phase)"
    x-axis ["Glass","Metal","Plastic","Wood"]
    y-axis "Accuracy (%)" 0 --> 100
    bar [86.4, 83.8, 89.5, 86.8]
    line [86.6, 86.6, 86.6, 86.6]

Loading

---

Classification Report

Class	Precision	Recall	F1-score	Support
Glass	0.90	0.86	0.88	879
Metal	0.84	0.84	0.84	756
Plastic	0.85	0.89	0.87	777
Wood	0.88	0.87	0.87	582
Accuracy			0.87	2994
Macro Avg	0.87	0.87	0.87	2994
Weighted Avg	0.87	0.87	0.87	2994

---