see-flowers

Flower Classification with Small Sample Data Using Deep Convolutional Neural Networks

Overview

This repo provides my work on flower image classification using deep convnet. I trained serveral models on VGG 102 flower dataset.
The best top-1 accuracy achieved so far: 0.904.
Note here that I follow the official protocol for dataset split.

• training: 1020
• validation: 1020
• test: 6149

Objectives

• Study the transfer learning on mulitclass, fine grained image classification task.
• Study the visualization tools and techniques for deep convnets.
• Try to explain the convnet-based classifier using visualization.

Report

see-flower

Usage

• Preparation

python init.py # download the dataset and organize

• Training

python train.py --model=[model_name]

Environment

python 2.7
keras 2
(to be completed...)

Models

Baseline model

A simple 2-layer baseline convnet.
acc: 0.312

VGG16

Fine tune VGG16 with weights pretrained on imagenet.
acc: 0.757

VGG19

Fine tune VGG19 with weights pretrained on imagenet.
acc: 0.781

Inception-v3

Fine tune inception-v3 with weights pretrained on imagenet.
acc: 0.904

Visualization

original

saliency

heatmap