How p-twinkle/perishpredict-smart-vision-for-produce-spoilage-detection works
• Data Source & Ingestion:
– Block for “Data Repository” indicating the dataset of ~30,000 images (fresh and spoilt produce) from Kaggle.
– “Data Preprocessing” component which handles image resizing, normalization, and splitting (training, testing, hold-out validation).
• Experimental Model Development:
– A module “Deep CNN Models” that further branches into subcomponents noted in notebooks (e.g., ResNet50, TensorFlowCNN, EfficientNet_binary).
– “Contrastive Learning Module” representing experiments with contrastive learning techniques.
– “SVM Classification Module” for the alternative machine learning approach.
– “LLM-Based Approaches” with subcomponents for “Zero-Shot” and “Few-Shot” experiments, representing the integration of vision–language models.
• Evaluation & Testing:
– An “Evaluation” or “Model Validation” component that listens to outputs from training modules and measures performance (accuracy between fresh and spoilt classification).
• Inference & Deployment (Future Direction):
– Even though deployment isn’t explicitly implemented, include a “Deployment/Inference Pipeline” that could eventually feed into quality control systems, smart refrigeration, or produce sorting mechanisms.
• Documentation & External References:
– Optionally, a component or annotation that indicates external sources (research papers, Medium posts, Kaggle datasets) which informed the design and experimental setups.
Relationships and Data Flow
– Start with the data source flowing into the “Data Preprocessing” block.
• Interactions such as model training feedback loops (e.g., performance metrics feeding back for hyperparameter tuning).
Architectural Patterns and Design Principles
• Separation of Concerns: Distinct blocks for data handling, model experimentation, and evaluation.
• Modularity: Each experiment (CNN variants, contrastive learning, SVM, LLM-based approaches) is shown as an independent module, making it easy to compare or extend.