When starting out, artists often find themselves struggling to properly convey their character's emotions. This project's purpose is to serve as a tool for begginer artists looking to make their emotions more captivating.

The dataset initially consists of 462 images distributed in the following manner:

• Pleased (38)
• Angry (54)
• Crying (56)
• Sad (57)
• Embarrassed (67)
• Happy (87)
• Shock/Surprised (103)

I added my own screenshots to the datasets, bumping it up to a total of 680 images distributed in the following manner:

• Pleased (80)
• Angry (86)
• Crying (82)
• Sad (81)
• Embarrassed (92)
• Happy (142)
• Shock/Surprised (117)

I used a pretrained ResNet-50, which I then finetuned with my custom dataset. I used Cross-Entropy loss as my loss function.

It was expected that the accuracy would not be satisfying given the size of the dataset. After multiple modifications, the model achieved 57.89% for both precision and recall. Below is the resulting confusion matrix for the testing dataset, which consists of 57 images.

• 0: angry

• 1: crying

• 2: embarassed

• 3: happy

• 4: pleased

• 5: sad

• 6: shock

    # Classes (rows)
    # Predictions (columns):
     0  1  2  3  4  5  6
    [5, 1, 0, 0, 0, 0, 0],
    [1, 3, 1, 1, 0, 1, 2],
    [0, 1, 3, 0, 0, 2, 0],
    [1, 1, 2, 4, 0, 0, 0],
    [0, 0, 0, 2, 5, 1, 1],
    [1, 1, 0, 0, 0, 5, 0],
    [0, 2, 2, 0, 0, 0, 8]]
  

Full-model loss VS Last layer loss

The model is deployed in a Flask web application (not deployed on the Internet). The user uploads an image file and clicks on upload. The model makes the prediction and the result is displayed.