Plant Disease Detection Using Mobile Images and Video

Mobile Image and Video Capture in the Field

Smartphones make plant disease detection accessible to farmers and gardeners because images can be captured directly in the field. Recent research shows that running deep learning models on mobile apps can help with early disease identification [2].

However, real-world photos taken outside are much harder to process than clean lab images. In a laboratory (like in parts of the PlantVillage dataset), leaves are typically picked, laid flat against a plain gray or black background, and photographed under even lighting. In a real field or garden, many environmental factors affect how leaves look.

Key Field Capture Factors

• Lighting and sunlight: Direct sunlight creates bright glare, washed-out spots, and dark shadows. This can change apparent leaf colors and make healthy leaves look damaged or hide real disease spots.
• Camera angle and leaf position: Leaves curl, droop, and overlap. A steep angle can distort the shape of spots or hide symptoms on the underside of the leaf.
• Distance from the leaf: If the photo is taken from too far away, small spots only cover a few pixels and are hard to detect. If taken too close, the overall leaf shape is lost.
• Background clutter: Soil, weeds, branches, and mulch can confuse models. The network might accidentally learn to associate soil type with disease instead of looking at the leaf itself.
• Motion blur: Hand shake and wind can blur fine textures like tiny fungal spots or leaf veins.
• Why video helps: A short video captures multiple frames from slightly different angles, helping overcome glare or blur that might ruin a single photo [5].

Image Preprocessing Steps

Preprocessing helps clean and standardize raw phone images before passing them to a neural network:

Common Preprocessing Techniques

• Resizing: Neural networks like YOLO and ResNet require fixed input sizes (such as 224×224 or 640×640). Using letterbox padding preserves the leaf's true aspect ratio without stretching it.
• Normalization: Pixel values from 0 to 255 are scaled to $[0.0, 1.0]$ and normalized using dataset channel means and standard deviations: $$\hat{X}_{c} = \frac{X_c - \mu_c}{\sigma_c}$$ This helps the model train and run more stably.
• Contrast enhancement (CLAHE): Contrast Limited Adaptive Histogram Equalization helps balance uneven lighting, making faint disease spots clearer.
• Color space conversion: Converting from RGB to HSV or CIE $L^*a^*b^*$ separates brightness from color, making it easier to spot yellowing (chlorosis) or browning (necrosis) regardless of lighting [3].
• Noise reduction: Filters like bilateral or median filtering reduce sensor noise while keeping leaf and lesion edges sharp.
• Data augmentation: Rotating, flipping, and slightly shifting colors during training helps models handle real-world variations without overfitting.

Leaf Segmentation and Background Removal

Removing soil, weeds, and neighboring branches allows downstream models to focus only on the leaf blade. While simple color rules (like excess green index $2G - R - B$) work in simple lighting, deep segmentation models (like U-Net) are much more reliable in complex outdoor scenes [3], [5].

Visual Disease Symptoms

Plant diseases produce visible changes:

• Chlorosis: Loss of green chlorophyll, showing as yellow patches or halos.
• Necrosis: Dead plant tissue that appears brown or black.
• Pustules and powdery patches: Spores from rusts or powdery mildew that create distinct surface textures.

Earlier computer vision systems measured these features by hand using color moments and texture formulas (GLCM) [3]. Modern deep learning models learn these visual patterns automatically from labeled training images.

Why Preprocessing Matters

Preprocessing cannot fix every photo, but it helps make input data more consistent. Proper sizing, normalization, and contrast adjustments help classification, detection, and segmentation models perform better in real-world use.