Machine Learning Project - Task: Images classification (17 Category Flower Dataset) . The data was obtained from the University of Oxford’s Department of Engineering Science.
You can download Dataset from given link : https://www.robots.ox.ac.uk/~vgg/data/flowers/17/
For more details and information about the experiments, you can refer to my PowerPoint presentation, which I created during the project. (I have placed my presentation, as well as the PDF and the results Excel sheet, in the ./Project_Analysis/Flower_Species_Classification.ppt directory.)
As we know that, having 17 classes, I have applies multiple algorithm (like : logistic regression , Single layer perceptron, multi layer perceptron, perceptron learning algorithm ect).
- Features are the information or list of numbers that are extracted from an image.
- When deciding about the features that could quantify plants and flowers, we could possibly think of Colour, Texture and Shape as the primary ones.
- In this Project I will be used three inbuilt libraries for
features extraction :
• Colour Histogram that quantifies colour of the flower
• Hu Moments that quantifies shape of the flower.
• Haralick Texture that quantifies texture of the flower.
• With MinMaxScaler normalization technique, its able to achieve 64.5% accuracy
• With StandardScaler normalization technique, its able to achieve 72% accuracy
• It’s performing best compared to all other algorithms that are evaluated.
• Train Accuracy : 88.83% & Test Accuracy : 60.41%
• For 5-fold it’s able to achieve an accuracy of 57% with my own code (No inbuilt)
• without k-fold ( train : 95% ; test : 45 %) with my own code (No inbuilt )
• For 5-fold it’s able to achieve an accuracy of 63.23% with 1 hidden layer of 20 hidden neurons and max iterations of 500 with my own code (No inbuilt )
• It can be observed that as the number of hidden neurons increases, the more linearity the trained dataset will tend to and increases the accuracy. Which is described by the cover’s theorem