Extracting Insights from Filters and Feature Maps

• Classification of leaf images such as Hypersensitive response (HR), normal, mosaic virus
• Visualization of filters and feature maps in leaf disease image classifiers
• In this way, we can visually identify which features have a significant impact on classification, and further extract the visual characteristics of the three kinds of leaf states.
• This repo is maintained by 오서영, 정명지
• Oct. 13, 2020

Dataset

• 804 images belonging to 3 classes by using google image crawling

Results

1. Baseline CNN with (32, 32) target size | Code

• 50 iterations, 1 batch
  Train accuracy : 97.60%
  Val accuracy : 96.63%

2. Baseline CNN with (128, 128) target size | Code

• 30 iterations, 1 batch
  Train accuracy : 98.40%
  Val accuracy : 97.19%

Visualization

0. Test sample - leaf infected with mosaic virus

1. Filters

• First Conv2D kernel size is (5,5) and second, third are (3,3)

2. Feature maps

• 3 feature maps with test sample (mosaic virus)