🍎 Fruit Quality Detector: Technical Architecture & Dual-Engine AI

---

📖 Introduction

The Fruit Quality Detector is a high-performance computer vision application. It solves the critical task of distinguishing between Good and Spoiled produce by analyzing visual features such as skin texture, color uniformity, and surface defects.

---

🧠 Core Concept: Convolutional Neural Networks (CNN)

What is a CNN?

A Convolutional Neural Network is a deep learning architecture inspired by the human visual cortex. Unlike standard neural networks that see images as a flat list of pixels, a CNN understands spatial hierarchies (edges → shapes → objects).

The 4 Key Layers in our CNN:

1. Convolutional Layer: Uses "filters" to scan the image. It acts like a magnifying glass looking for specific patterns (like a brown spot on an apple).
2. Activation (ReLU): Adds non-linearity. It decides which features are "important" enough to pass to the next layer.
3. Pooling (Max Pooling): Reduces the image size while keeping the most important information. This makes the model faster and more robust to image rotation.
4. Dense (Fully Connected): The final "brain" that takes all the extracted features and makes the final decision: Is this a good orange or a spoiled one?

---

🏛️ Custom Model: MobileNetV2 Backbone

What is a "Backbone"?

In deep learning, a Backbone is a pre-trained model that acts as a "Feature Extractor." We use MobileNetV2 as our backbone. It has already "seen" millions of images (ImageNet) and knows how to recognize shapes, colors, and textures.

• Why MobileNetV2?: It is designed for speed. It uses Depthwise Separable Convolutions to provide high accuracy while using very little memory.
• The Custom Head: We removed the original classification layer of MobileNetV2 and added our own Custom Layers (Dense, Dropout, Softmax) to specifically detect fruit quality.

---

🧬 ResNet-50: Residual Architecture

Advanced Architecture Breakdown

ResNet (Residual Network) is famous for its Skip Connections.

Standard networks try to learn the full mapping $H(x)$. ResNet instead learns the "Residual" $F(x) = H(x) - x$. This allows the network to effectively "bypass" layers if they aren't helping, which prevents the Vanishing Gradient Problem (where the model stops learning because it's too deep).

ResNet-50 Data Flow Diagram

graph TD
    subgraph "Phase 1: Stem"
    In["Input (224x224x3)"] --> C1["7x7 Conv (64 Filters)"]
    C1 --> BN1["Batch Norm + ReLU"]
    BN1 --> MP1["3x3 Max Pool"]
    end

    subgraph "Phase 2: Residual Stages"
    MP1 --> S1["Stage 1: 3x Bottleneck Blocks"]
    S1 --> S2["Stage 2: 4x Bottleneck Blocks"]
    S2 --> S3["Stage 3: 6x Bottleneck Blocks"]
    S3 --> S4["Stage 4: 3x Bottleneck Blocks"]
    end

    subgraph "Phase 3: Classification Head"
    S4 --> GAP["Global Average Pooling"]
    GAP --> FC["Fully Connected (1000 units)"]
    FC --> SM["Softmax (Probabilities)"]
    end

    style In fill:#f9f9f9,stroke:#333
    style SM fill:#dcedc8,stroke:#33691e,stroke-width:2px

Loading

---

📊 Models Table: Side-by-Side

Feature	Custom CNN (MobileNetV2)	ResNet-50
Logic	Specialized Fine-Tuning	Pre-trained Generalist
Architecture	Depthwise Separable Conv	Residual Skip Connections
Primary Goal	Fruit Freshness	General Object Identity
Performance	High Accuracy on this Dataset	Baseline Comparisons

---

🚀 Setup & Execution

Note

Ensure you have TensorFlow 2.15.0 installed for maximum compatibility with the .h5 model files.

# Install dependencies
pip install -r requirements.txt

# Start the Streamlit server
streamlit run app.py

---

Developed for professional fruit quality assessment using Deep Learning.