imageomics/butterfly_segmentation_unet

Model Card for butterfly_segmentation_unet

This model takes in an image of a butterfly (with or without body attached to wings) and segments out any existing hindwings and forewings, in addition to pictured equipment described below.

Model Details

unet_butterflies_256_256.hdf5 is the butterfly segmentation model.

The segmentation model was trained on a dataset of 800 total images from the Jiggins, OM_STRI, and Monteiro datasets. The model architecture is based on a simple UNet architecture for multiclass classification.

Model Description

Keras implementation of Butterfly UNet segmentation model. The model is responsible for taking an input image (256 x 256 x 3) and generating segmentation masks for all classes below that are found in the image. Vertical flips were applied as data augmentations prior to training the model. Images are converted to grayscale before being fed into the model. The model is trained with categorical cross entropy loss.

Segmentation Classes

[pixel class] corresponding category

• [0] background
• [1] generic
• [2] right_forewing
• [3] left_forewing
• [4] right_hindwing
• [5] left_hindwing
• [6] ruler
• [7] white_balance
• [8] label
• [9] color_card
• [19] body

Class weights

The model was fed class weights to help improve performance on wing segmentation categories (pixel classes: 2,3,4,5). The class weights dictionary used is:

{0: 1.0, 1: 1.0, 2: 3.0, 3: 3.0, 4: 3.0, 5: 3.0, 6: 1.0, 7: 1.0, 8: 1.0, 9: 1.0, 10: 1.0}

Developed by: Michelle Ramirez

How to Get Started with the Model

To view applications of how to load in the model file and predict masks on images, please refer to this github repository