Local Crop Classifier Study
Agricultural deep learning models trained on perfect, uniform datasets fail catastrophically in real-world farm conditions where lighting, shadows, and occlusions are highly unpredictable.
The Solution
I manually captured and preprocessed 1,200+ photos of regional vegetables under varying wholesale market lighting conditions. I then compared six different convolutional network designs, evaluating lightweight depthwise-separable designs against heavier ResNet configurations. To ensure robustness, I built an OpenCV-based background segmenter and class-balancing filter to isolate organic structures in real-time.
- Manually captured and preprocessed over 1,200 photos of regional vegetables (Karela, Parwal) in varying wholesale market lighting conditions
- Evaluated accuracy and hardware latency trade-offs across 6 networks, including depthwise-separable designs
- Wrote custom OpenCV filters to perform automatic background isolation and HSV color segmentation
- Configured a Flask web interface to test real-time validation speeds and latency metrics
System Pipeline
Engineering Trade-offs
Why?I traded absolute top-1 accuracy on pristine images for robust performance on noisy images by employing synthetic minority oversampling (SMOTE) and aggressive affine transformations. I chose depthwise-separable convolutions to guarantee the model could run on low-end hardware, even though it meant sacrificing some edge-case detection capabilities.
Proven Impact
Identified optimal network parameters that achieved 92% real-world accuracy while reducing inference latency by 45% compared to default ResNet configurations, making it viable for low-cost edge deployments.