Three sets of images have been provided for the Skin Recognition project:

1. Training images
2. Skin mask of the training images
3. Test set Each training image and its mask have the same size. For each pixel in the original image there is a corresponding pixel in the skin mask with. The corresponding pixel in the mask is not skin the pixel has the RGB (1, 1, 1); otherwise the skin pixel is present.

The SkinRecognition class has the following attributes:

• imgpath: Path to images (for training or testing).
• maskpath: Path to skin mask images.
• rect: The MxN neighborhood grid.
  • default: (1, 1). This means only the pixel itself is considered as a feature.
• knn_neighborhoods: K Nearest Neighbors to be used in the KNN algorithm.
  • default: 7
• img_paths: list of images provided by imgpath.
• mask_paths: list of masks provided by maskpath.
• X_training: Features of input images.
• Y_training: Labels of the features.
• knn: The KNeighborsClassifier object.
• confusion_matrix: If in test method generate_confusion_matrix is set, this attribute holds the average confusion matrix over all the tested images. Each confusion matrix is output to a csv file.

For this the method get_training_images is used. The input parameters are listed below:

• use_kmeans_reduction: Whether to use kmeans reduction technique or not. This way after reading each set of image, the number of output features are reduced by a specified factor.
  • default: True
• skin_clusters: The K value of the Kmeans algorithm. The number of output centroids. if the value is between 0 and 1, the number of clusters would be calculated as skin_clusters×length(features). It is ignored if use_kmeans_reduction is set to False.
• non_skin_clusters: The same logic as for skin_clusters applied to non_skin_clusters.
• knn_neighborhoods: K Nearest Neighbors to be used in the KNN algorithm
  • default: None, which uses the same value as provided in the initializer.
• img_start_index: The index where training would start on the list of images obtained from imgpath.
• img_count: Number of images to be read in the training.
• slice_size: How many images to be read in each iteration.
  • default: 3
  • This is specially useful in two scenarios:
    • When output model is set to a file so after each iteration over the slice the features are stored in the disk,
    • When using kmeans reduction, the numbers of skin and non-skin clusters are applied to the iteration
• output_model: The path and filename where the output model should be stored.
  • default: None. The output model will not be stored in disk.
• image_reduction_factor: The reduction factor to be applied on the image.
  • default: 0. No reduction is applied on the image.

sk = SkinRecognition(imgpath='PATH/TO/IMAGES', maskpath='PATH/TO/MASKS', rect=(7, 7), 
  knn_neighborhoods=8, model='PATH/TO/MODEL')
sk.test(img_test_count=10, img_count=0, method='all', generate_confusion_matrix=True, 
  output_results='PATH/TO/RESULTS', output_mask_path='PATH/TO/MASKS')

The parameters for this method are:

• test_count: The number of images from the list obtained from imgpath to be tested.
• test_start_index: The first image from the list obtained from imgpath should be the starting point.
• method: 'rect' or 'all'
  • default: 'rect': Only tests over the MxN grids. This method is faster but does not create neighborhoods for each pixel and predicts the class over each grid.
  • 'all': Creates MxN neighborhood for each pixel and predicts the class one by one.
• generate_confusion_matrix: Whether to cross validate the image if the mask for it exists. It iterates through each predicted mask and compares it against the existing mask.
  • default: True. It can be done for cross validation when masks exist.
• output_results: Path to output result file which is in CSV format. It includes the path to the file, the time needed for test, and the confusion matrix of the prediction.
  • default: None. Does not generate output results.
• output_mask_path: The path to store the masks.
  • default: None. Does not generate the predicted mask image file. Instead shows the results on screen using the matplotlib library.