Hyperparameter for grascale image training.

[harufumigithub]

I am working on a project to classify biomedical conditions from a set of images. Although the input images are all in grayscale, they possess distinct shape features. Originally, these images were captured in RGB but were converted to grayscale to simplify the problem. Our approach involves using segmentation to train the model on these images.

However, we are encountering a challenge where the model exhibits bias towards specific shades of grayscale. It appears that objects with identical features but varying colors (in their original RGB format) are not recognized as belonging to the same class after conversion to grayscale. This issue persists despite adjustments to various hyperparameters.

We suspect that the model's bias might be related to how the grayscale conversion process affects the visibility of certain features, particularly for objects originally in red hues. To address this, we have experimented with augmenting the images in the HSV (Hue, Saturation, Value) color space, but the outcomes have not been satisfactory. Below are the HSV augmentation parameters we have utilized:

hsv_h: 0.0 (float) - Image HSV-Hue augmentation (fraction)
hsv_s: 0.7 (float) - Image HSV-Saturation augmentation (fraction)
hsv_v: 0.4 (float) - Image HSV-Value augmentation (fraction)

Given this context, I am seeking recommendations on which hyperparameters might be effectively adjusted to mitigate this bias, along with reasonable values for these parameters. Any insights or suggestions from your experiences would be immensely valuable. We plan to try out hsv_s: 0.0 and hsv_v: 100 but would like your input first before committing lengthy training period.

Thank you in advance for your assistance.

[glenn-jocher]

@harufumigithub the issue you're facing is not uncommon when converting RGB images to grayscale for classification tasks. The conversion process can obscure certain features that were originally distinguishable by color, leading to a loss of information that can be critical for accurate classification.

Here are some suggestions to address the bias and improve your model's performance:

1. Reconsider Grayscale Conversion: If color information is important for distinguishing between classes, consider using the original RGB images for training. Modern deep learning models can handle the complexity of color images and may perform better with the additional information.

2. Custom Grayscale Conversion: Instead of using a standard grayscale conversion, you could experiment with custom conversion formulas that weigh the RGB channels differently. This could help in preserving the distinguishing features of objects that are not well-represented in a standard grayscale image. For example, you could increase the weight of the red channel if red features are important.

3. Data Augmentation in RGB: Before converting to grayscale, you could augment the images in the RGB space. This might include adjusting brightness, contrast, and performing color jittering (randomly changing the brightness, contrast, saturation, and hue of an image). This can help the model generalize better across different shades and colors.

4. HSV Augmentation Parameters: The parameters you've chosen for HSV augmentation seem to be quite aggressive, especially for saturation (hsv_s: 0.7). This could lead to unnatural images that do not represent the true variation in your dataset. You might want to try more subtle augmentations, such as hsv_s: 0.1 to 0.3 and hsv_v: 0.1 to 0.3, to start with. Also, consider using random shifts in hue (hsv_h) to simulate different color variations.

5. Class Balancing: Ensure that your dataset has a balanced representation of the classes across different shades of grayscale. If certain shades are underrepresented, the model may develop a bias towards the more common shades.

6. Feature Engineering: If certain shapes are critical for classification, consider using shape descriptors or other feature extraction techniques that can capture shape information independently of color.

7. Model Architecture: Some neural network architectures might be more sensitive to color and texture features than others. Experiment with different architectures to find one that captures the relevant features for your task.

8. Loss Function: If your model is exhibiting bias, you might want to look into using a loss function that penalizes the model more for misclassifying underrepresented classes. For example, focal loss is designed to focus more on hard-to-classify examples.

9. Post-processing: After training, you could apply post-processing techniques to adjust the model's predictions based on known biases. For example, if you know that the model is less accurate for certain shades, you could apply a correction factor to the predictions for those shades.

10. Evaluation Metrics: Use evaluation metrics that account for class imbalances and biases, such as the F1 score, Matthews correlation coefficient, or weighted accuracy.

Before committing to lengthy training periods, it's advisable to perform quick experiments with a subset of your data to validate the effectiveness of these strategies. This iterative approach can help you identify the most promising solutions without investing excessive time and resources.