AI-assisted crop leaf disease detection.
CropCheckUp analyzes a leaf image, removes the background, prepares the image for a TensorFlow Lite model, and shows the predicted crop condition with confidence and management information.
The app is intentionally narrow: one leaf image goes through segmentation, model-ready preprocessing, TensorFlow Lite inference, then a readable diagnosis summary.
The user captures a live camera frame or selects an image from the gallery.
CropCheckUp starts with one visible leaf. A clean single-subject image gives the downstream segmentation and classifier less irrelevant texture to absorb.
The app isolates the leaf so the classifier focuses on the plant area.
Background removal reduces noise from soil, hands, pots, shadows, and neighboring plants before the model sees the image.
The processed image is resized to 224 x 224 pixels.
Every image is normalized to the same square input size, matching the TensorFlow Lite model shape used during training and export.
A TensorFlow Lite model predicts the most likely crop condition.
The app converts the resized RGB image into a tensor, runs local inference, and reads the top-scoring label from the output vector.
The app displays crop name, condition, confidence, symptoms, causes, and management guidance.
Raw model labels are translated into readable names and paired with the supporting context a user needs before deciding what to inspect next.
The landing page exposes the exact assumptions that matter for this app: input size, color channel handling, label conversion, and what happens to transparent background pixels.
Raw leaf photo enters from the mobile capture flow.
Leaf pixels are isolated so the model sees less surrounding scene noise.
The isolated image is resized to the model's expected square input.
Pixels are packed as RGB values before inference.
The on-device model returns a score vector across the label file.
The top label is converted into crop, condition, confidence, and guidance.
The model receives a 224 x 224 RGB image. Pixel values are passed in the 0-255 range because preprocessing is included inside the model.
Transparent background pixels are treated as black during tensor conversion.
The highest scoring label is resolved against the app's label file before the result screen is rendered.
The classifier returns one top label and a score. The app turns that raw model output into a result a grower can scan quickly: crop, condition, confidence, symptoms, causes, management guidance, and the processed leaf preview.
The classifier returns the highest-scoring label from the model output. Labels follow a crop and condition format such as Crop___Disease or Crop___healthy.
The app converts raw labels into readable crop and disease names, keeps the confidence percentage visible, and separates advisory content from the model prediction.
Label file: The current label file contains 68 crop and condition labels.
The model development workflow is documented in the Kaggle notebook. The dataset used for CropCheckUp experiments and training is available on Kaggle.
Model development workflow and experimentation.
Image dataset used for CropCheckUp experiments and training.
CropCheckUp derived dataset and bundled model artifacts are licensed CC-BY-NC-SA-4.0 for non-commercial use.
CropCheckUp is an AI-assisted screening tool. Results depend on image quality, lighting, leaf visibility, and whether the condition exists in the trained labels.
Use a clear image of one leaf. Keep the leaf in focus, fill the frame with the relevant surface, and avoid mixing multiple plants in one capture.
Avoid blur, shadows, and cluttered backgrounds. Harsh light, dark shadows, soil texture, hands, and pots can make segmentation and classification less reliable.
The app can only classify supported labels. Unknown crops, new diseases, mixed symptoms, pests, nutrient stress, or water damage may still map to the nearest known label.
