This is a computer vision + machine learning project trained using the FastAI resnet152 CNN model to classify 54 musical instruments based on images.

This project demonstrates the use of deep learning to classify a wide range of musical instruments from images with strong and scalable accuracy.

• Recognizes 54 musical instruments across 6 categories (stringed, keyboard, brass, woodwind, percussion, and miscellaneous instruments).
• Data collected (~1000 images per instrument) using DuckDuckGo Search API.
• Trained with resnet152 and fine-tuned over 10 epochs of augmented data.
• Interactive prediction interface via image upload in Colab.
• Includes top-3 prediction chart and natural language output.

• musical_instrument_classifier.ipynb – Google Colab notebook.
• musical_instrument_classifier_resnet18.pkl – Trained model (v0.5.a).
• musical_instrument_classifier_resnet152.pkl – Trained model (v1.0.a).
• confusion_matrix_resnet18.png – Class performance visualization of the model pre-trained with resnet18.
• confusion_matrix_resnet152_v1.png – Class performance visualization of the model pre-trained with resnet152.
• top_losses_resnet152.png – Top losses list of the model pre-trained with resnet152.

• Python
• fastai / PyTorch
• resnet152
• DuckDuckGo API (for image scraping)
• Hashlib (for duplicate scraped images)
• Pathlib and PIL (for corrupt/unsupported image file handling)
• Matplotlib / ipywidgets
• Google Colab

• Unconventional or niche instruments may produce inaccurate predictions.
• Visually complex instruments may confuse the model or provide accurate but relatively unconfident results.
• The accuracy and generalization performance of the model suffer due to the intrinsic disadvantage of having a general dataset.

• Expand the training dataset strategically by providing special queries highlighting the qualities specific to each instrument. (v0.5.b)
• Add more instrument types, non-standard angles, and backgrounds. (v1.0.b)
• Include Gradio or HuggingFace for the model's UI application. (v1.0)

• MIC model (v1.0.a) trained using resnet152. Data cleaning and augmentation were performed using approximately 1000 training data points per instrument.
• MIC model (v0.5.a) trained using resnet18. No data cleaning and augmentation with imbalanced and limited (~300-400) training data points per instrument.

• This project is licensed under the MIT License.