Cigarette Detecting and Blurring Using Mask RCNN

This is an image segmentation project which detects and extracts the masks of cigarette instances in images using Mask R-CNN. The Mask R-CNN model generates bounding boxes and segmentation masks for each instance of an object in the image as the paper suggests.

• Data
  • To access the unlabeled images
  • To access the processed and labeled images
  • Data Preprocessing
• Results

Data

There are 731 unlabeled images of cigarettes scraped only from google images and 193 of them are labeled pixel-by-pixel. Further scraping will be done from stock websites.

Since the dataset is stored in my personal Google Drive account and the size of the dataset is larger than 100MB, I had to use a script in this reposityory of the user @circulosmeos which directly downloads big files from Google Drive sharing links.

To access the unlabeled images:

Download the gdown.pl script

wget -O gdown.pl "https://drive.google.com/uc?export=download&id=1gc-D4ydSBY7jqhlM-Oq2xca5oJx9217B"

Permit the user to run the script

chmod +x ./gdown.pl 

Download the data using the script

./gdown.pl 'https://drive.google.com/file/d/1-6QvZMUB13r-bN1eoiF3qlEplS_4imiK/view?usp=sharing' 'data-cigar-731.zip'

Unzip it

unzip ./data-cigar-731.zip -d /path/to/extract

To access the processed and labeled images:

./gdown.pl 'https://drive.google.com/file/d/1nHKQMUbhNBAyn307K5jEJmX2PBK7rlfz/view?usp=sharing' 'cigarette.zip'

Data Preprocessing

Because the dimensions of the images must be 1024x1024 before using Mask-RCNN model;

• images smaller than the target size got expanded using np.pad() function of Numpy module

• images larger than the tarhet size got resized using resize() function of OpenCV module

Training

You need to clone the Mask-RCNN GitHub Repository and configure the config files/codes according to your workspace and data. Instructions are given in the repository.

Results

Despite having a very shallow dataset, using transfer learning leads the model having some baseline results.

Since the model has been trained with only ~200 images, it easily gives false results in some cases.