This project leverages deep learning to classify children's psychological health based on their drawings. Using pretrained models such as EfficientNetB0, EfficientNetB1, MobileNetV2, ResNet50, and YOLOv8n-cls, the model was trained and evaluated to determine the best approach for real-time classification. The top-performing model, YOLOv8n-cls, is designed to be deployed in web or mobile applications to help parents monitor and analyze their children's drawings, potentially helping identify signs of psychological stress or well-being.
This work was deeply inspired by and built upon the findings from the paper:
"A Children's Psychological and Mental Health Detection Model by Drawing Analysis based on Computer Vision and Deep Learning" by Amal Alshahrani, Manar Mohammed Almatrafi, Jenan Ibrahim Mustafa, Layan Saad Albaqami, and Raneem Abdulrahman Aljabri, available here.
Their research provided valuable insights into the application of the YOLOv8n-cls model, which demonstrated a top-1 accuracy of 94% at epoch 10 with a compact model size, making it highly efficient for mobile applications. Motivated by their work, I retrained their dataset to achieve comparable results, reaching a top-1 accuracy of 95.99% at epoch 10.
The dataset used for training and evaluation was collected from children aged 6 to 15 in Mecca. It consists of 500 images, split as follows:
- Training set: 80% of the images
- Validation set: 10%
- Test set: 10%
The dataset is labeled into four categories:
- Anger and aggression
- Anxiety
- Happy
- Sad
You can access the dataset here.
The dataset preparation process involves organizing raw data into class-specific folders, making it ready for training. This step is automated using the prepare-dataset.ipynb notebook in the repository.
The evaluation of the pretrained models showed the following results:
Note: YOLOv8n-cls achieved the highest performance and was selected as the best model for real-time analysis. The deployment code for the YOLOv8n-cls model is hosted on Render and is available to test.
The next step involves integrating the trained model into a web or mobile application, enabling parents to access real-time analysis of their children's drawings and store them for ongoing monitoring.