Depression Detection from Visual Content

Key Highlights

• Developed a binary classification system to detect depression indicators from image data using both traditional ML and deep learning approaches on a 20,186-image dataset curated from social media visual content.
• Implemented Logistic Regression with grayscale image features (32×32) achieving 75% accuracy with strong generalization; built a ResNet18 transfer learning model with ImageNet pre-training for comparison.
• Analyzed precision, recall, F1 scores, confusion matrices, and computational efficiency across both architectures; provided deployment recommendations based on accuracy/speed/resource trade-offs.
• Performed extensive preprocessing including resizing, normalization, and data augmentation; addressed class imbalance via stratified sampling and weighted cross-entropy loss.
• Demonstrated ResNet18 outperformed Logistic Regression by ~8% F1 on the held-out test set while requiring significantly more compute — highlighting trade-offs relevant for edge deployment.
• Documented all findings in a structured experimental report with an ablation study comparing feature engineering strategies, batch sizes, and learning rate schedules.