Dylan Pilz
Last updated: 7/9/2021
The CIELAB colorspace is of great importance in food image evaluation since it mirrors human color perception more accurately than the typical RGB colorspace used by most cameras. It is therefore better suited for predicting a consumer's perception of food quality.
In this project, I used several approaches to attempt to recreate the results of León et al, which seeks an accurate approach to convert pixel values from RGB to LAB for use in food evaluation.
The data is taken from an image of a standard color checker, with known Lab* values.
Linear regression approach, using linear_model from sklearn library.
subset_data(rgb_file, LAB_file) Clean input .txt files and return them as numpy arrays.
plot_regression(X, y, title) Produce scatter plot and best fit line
calculate_stats(expected, predicted, title) Prints summary statistics(shown below) and returns mean absolute error
Quadratic regression, again using linear_model but with RG, RB, GB, R^2, G^2, B^2 appended to input array
subset_data(rgb_file, LAB_file) Clean input .txt files and return them as numpy arrays.
quad_setup(rgb_data) Additional preprocessing step, appending RG, RB, GB, R^2, G^2, B^2 to input array
plot_regression(X, y, title) Produce scatter plot and best fit line
calculate_stats(expected, predicted, title) Provides summary statistics(shown below) and returns mean absolute error
quad_setup(rgb_data)
Splits LAB input into seperate columns, and trains a seperate model for each, with 2 hidden layers containing 4 nodes each. Performs decently well, but could be improved significantly given more training data. The accuracy for L values tapers off at the extremes, but this shouldn't be too much of an issue when evaluating meat since the lightness values remain fairly centered around 50.
subset_data(rgb_file, LAB_file, col) Parse input files and output numpy array, the col parameter specifies which column (L, a, or b) to select from the LAB_file. Scales L* values down by a factor of 100, a* and b* values by 128
build_model(output_activation) Construct a model with the provided output activation function (sigmoid for L*, tanh for a* and b*) Contains 2 hidden dense layers with 4 nodes each.
plot_loss(hist, label) plots the output of the loss function for each epoch
Mean absolute errors for the three models are shown below:
| Value\Model | Neural network | Linear regression | Quadratic regression |
|---|---|---|---|
| L* | 1.8639297685169036 | 4.5966228965582285 | 2.187636465294087 |
| a* | 6.498623516630558 | 7.017837540294636 | 3.0173261809247562 |
| b* | 2.848936311006546 | 4.7359271862145285 | 2.4048787362640494 |
Katherine León, Domingo Mery, Franco Pedreschi, Jorge León, Color measurement in L∗a∗b∗ units from RGB digital images, Food Research International, Volume 39, Issue 10, 2006, Pages 1084-1091, ISSN 0963-9969, https://doi.org/10.1016/j.foodres.2006.03.006.