Poison Ivy Detection via Semi-supervised Image Recognition

Overview

This project tackles the problem of poison ivy detection via a semi-supervised image recognition algorithm in MATLAB. The aim of the project is to inform users if there is a poison ivy in a given picture. We used a combination of feature detection methods like CNN, corners and blob detection for classification.

Methodology

Pre-processing: Initial Morphology and Noise Reduction

• First, the algorithm finds the ratio of the image dimensions and explores the colorspace for high-signal channels.

• A circular mask is generated to remove the area outside the circle.

• Image quantization is used to bucket the pixels of the image into 16 parts to further remove noise.

Processing: Strong Edges

• Sobel edge detector is used to detect edges in vertical and horizontal directions.
• Edges with a magnitude above the 95th percentile are identified to achieve the strongest edges.

Processing: K-Means and Noise Removal

• K-means is used for segmentation, multiple color channels are inputted into K-means.
• 5% weight for rows and columns, 200% weight for edge magnitudes, and a* from Lab are used for adjustments.
• Clusters for K-means are set to nine.

Processing: Watershed Segmentation

• If K-means finds a single connected component of leaves, separate leafs with watershed segmentation

Post-processing: Cleaning and Classifying

• Morphological operations are used to clean the image and prep for classifying algorithms to take over
• Classifying factors are: Number of leafs, number of corners, and neural network's output on the post-processed leaf image

How to use

Enter the src file for details on how to use. Read the "Final Report.pdf" for a more detailed explanation of the methodologies.

Results

Our algorithm has been tested on a set of images and was able to detect poison ivy with an accuracy of ~94%.

Contributors

• Bardh Rushiti
• Mohd Junaid Shaikh
• Saksham Bansal