Fully Configurable N-Layer Neural Network

Overview

A configurable multi-layer neural network implementation for image classification tasks, featuring both training and inference capabilities with backpropagation optimization. One main application/configuration detects the number of fingers (1-5) someone is holding up in hand gesture images.

Key Features

• Finger Recognition: Classifies hand gestures into 5 categories (1-5 fingers)
• Training Mode: Trains from scratch with random weights on 25 training cases
• Inference Mode: Uses trained weights to classify new images (6th set of test cases)
• Configurable Architecture: 13,000 input nodes → 25 → 5 → 5 output nodes
• Backpropagation Training: Optimizes weights using gradient descent

Files Structure

Core Implementation

• NLayer.java - Main neural network implementation

Configuration Files

• ImageProcessingConfig.txt - Main network configuration
• config2N1.txt - Alternative 2-1-1 network configuration
• config3N3.txt - Alternative 2-1-1-3 network configuration

Training Data

• TestCases/ - Directory containing finger gesture image data
  • {1-5}-{1-5}.txt - Training cases: 5 sets of 5 images each for fingers 1-5
  • 6-{1-5}.txt - Test cases: 6th set used for inference mode
  • Each file contains 13,000 normalized pixel values (120×100 + padding)
• weights.txt - Saved neural network weights after training

Image Preprocessing (To Test Your Own Images)

• WOOHOO.py - Python script for image preprocessing
  • Converts images to grayscale
  • Removes background using AI-based segmentation
  • Centers and crops images to fixed size
  • Outputs 120×100 pixel images as BMP format

Training vs Inference Modes

Training Mode (Default)

• Purpose: Train the network to recognize finger counts
• Data Used: 25 training cases (sets 1-5, all 5 images per finger count)
• Configuration:
  • willTrain = true
  • useRandomWeights = true
  • useLoadedWeights = false
• Output: Saves optimized weights to weights.txt

Inference Mode

• Purpose: Use trained network to classify new finger images
• Data Used: 5 test cases from set 6 (6-1.txt through 6-5.txt)
• Configuration Changes Required:
  • willTrain = false
  • useRandomWeights = false
  • useLoadedWeights = true

Usage

Quick Start

Step 1: Train the Network (Default)

java NLayer

By default, the system trains on 25 finger gesture images with random weights until the error threshold is reached, then saves weights to weights.txt.

Step 2: Switch to Inference Mode

To use the trained network to classify new finger images:

1. Edit ImageProcessingConfig.txt:

   • Change willTrain = false
   • Change useRandomWeights = false
   • Change useLoadedWeights = true

2. Run inference:

java NLayer

This will test the network on the 6th set of finger images and show classification results.

You can configure the network for any other tasks as you'd like.

Output Classification

The network classifies finger gestures into 5 categories:

• 1 Finger: Cases 0-4 (files 1-*.txt) → Output [1,0,0,0,0]
• 2 Fingers: Cases 5-9 (files 2-*.txt) → Output [0,1,0,0,0]
• 3 Fingers: Cases 10-14 (files 3-*.txt) → Output [0,0,1,0,0]
• 4 Fingers: Cases 15-19 (files 4-*.txt) → Output [0,0,0,1,0]
• 5 Fingers: Cases 20-24 (files 5-*.txt) → Output [0,0,0,0,1]

Dependencies

Java

• JDK 8 or higher
• Standard Java libraries (Properties, IO)

Python (optional)

• NumPy
• PIL (Python Imaging Library)
• SciPy
• rembg (for background removal)

Author

Anish Jain (Version 4.30.24)