cnizinski / image_textures

Functions for extracting image texture features by angle measurement technique (AMT), gray level co-occurrence matrices (GLCM), and convolutional neural networks (CNN).

image_textures

• Angle measurement technique (AMT)
  • A spectral texture analysis method that measures mean angles along unfolded grayscale image spectra at different scales
  • The mean angle AMT spectra reveal information about the image's compexity and periodicity on certain scales
• Gray level co-occurrence matrix (GLCM)
  • A statistical texture method that computes directionally-independent features from co-occurrence matrices of certain quantization levels and distances.
• Convolutional neural networks (CNNs)
  • Automatic feature extraction using the output of the next to last layer.

Modules and Functions

helpers.py

Functions for data handling and various mathematical functions.

• interpolate(xi, p0, p1) - Linear interpolation between 2 points
• img_info(fname, fields) - Returns dictionary of image info from regularly formatted filenames; the dicts from each image can quickly be converted to a pandas dataframe for data exploration and filtering
• quick_filter(df, filt_dict) - Returns filtered dataframe from info in dictionary
• json2df(dpath, dfiles) - Returns dataframe from set of json data files

preprocessing.py

Functions for sampling and cropping image data.

• sample_img(img, barh, rows, cols, region) - Samples region of SEM image given grid size and region index
• random_crop(img, random_crop_size) - Randomly samples a square region SEM image with specified size

amt.py

Functions needed to implement angle measurement technique (AMT) texture analysis.

• unfold(img, snake) - Unfolds grayscale image into 1-D spectrum, "snake" unfolding goes left-to-right then back right-to-left
• get_lr(imgspec, cpt, scale) - (OBSOLETE and SLOW, use get_left + get_right instead) Gets left and right points for angle from center point and distance/scale
• get_left(imgspec, cidx, scale) - Returns left point for angle measurement from center index and scale
• get_right(imgspec, cidx, scale) - Returns right point for angle measurement from center index and scale
• calc_angle(cpt, lpt, rpt) - calculates angle from center, left, and right points using dot product
• img_amt(img_roi, max_scale, n, snakes) - Runs AMT for single image region up to max_scale index and with n points sampled, returns dictionary of mean angles
• batch_amt(img_df, img_path, save_path, label, params) - Runs AMT for batch of images (whose filenames are listed in the dataframe), returns dict with each image's AMT data and saves it as json file

glcm.py

Functions needed to implement gray level co-occurrence matrix (GLCM) texture analysis and to compute Haralick texture features.

• quant_img(img, q) - Returns quantized image with q bins from grayscale image
• get_glcms(img, levels, dist) - Returns co-occurrence matrices at 0, 45, 90, and 135 degrees at specified distance and with specified gray level quantization
• Haralick features and helper functions - functions for computing Haralick features from the 4 directional GLCMs
• glcm_features(glcms_dict, sf) - Produces dictionary for each Haralick feature from 4 directional GLCMs, allows for dict key suffix to be specified
• batch_glcm(img_df, img_path, save_path, label, params) - Produces GLCM feature data for a batch of images (whose filenames are listed in the dataframe), saves to file and returns dict for batch

cnnfeatures.py

Uses pre-trained VGG19 classifier to extract output of FC layer before predictions.

• batch_vgg19(img_df, img_path, save_path, label, params) - Returns FC output and ImageNet predicted label for a batch of images (as a pandas dataframe)