Shubham12864/plant-leaf-disease-detection-using-cnn

🌿 Plant Leaf Disease Detection Using CNN

Official Capstone Project & Benchmark: plant-leaf-disease-detection-using-cnn
Author: Shubham Kumar (Shubham12864)
Deployment Formats: PyTorch (.pt), ONNX FP32 (.onnx), and Edge-Optimized Dynamic INT8 (.onnx)

---

🔬 Architectural Innovations

Eliminates the limitations of discrete multi-tool pipelines (YOLO cropping + classifier glue code) by creating a single, unified, end-to-end differentiable neural network:

Input Image (224x224x3)
         │
         ▼
[ConvNeXt-V2 Backbone with GRN]
Stage 0 (96-d) ──► Stage 1 (192-d) ──► Stage 2 (384-d, 14x14)
                                              │
                                              ▼
                         [Differentiable Spatial Saliency Gating]
                         M(x) in [0, 1] via GroupNorm(1, C) + Conv2D
                         Controlled Attenuation: F_gated = F * (0.1 + 0.9 * M)
                         (Autonomous 90% non-foliar clutter suppression)
                                              │
                                              ▼
                                       Stage 3 (768-d)
                                              │
                                              ▼
                         [Anatomical Hierarchical Dual Head]
                         ├── Head 1: Crop Family (k=14, 91.04% Acc)
                         └── Head 2: Conditioned Pathology (k=38, 85.59% Acc)
                                              │
                                              ▼
                         [Free Helmholtz Energy Safety Head]
                         E(x) = -T * logsumexp(logits / T)

1. Autonomous Spatial Gating Module: Learns to distinguish healthy/diseased leaf foliage from non-foliar background (soil, hands, pots, sky) directly in latent feature space: $$F_{\text{gated}} = F \odot \Big(\alpha + (1 - \alpha) \cdot M(x)\Big), \quad \alpha = 0.1$$ Attenuates background clutter by $90%$ without requiring external bounding-box object detectors.
2. Hierarchical Taxonomic Conditioning: Conditions disease diagnostic logits on soft crop family probabilities, achieving $93.83%$ biological consistency and preventing cross-species hallucinations.
3. Split Conformal Prediction Safety Guard: Provides mathematically guaranteed prediction sets $\mathcal{C}(X)$ achieving $96.61%$ empirical field coverage at $1 - \alpha = 95.0%$ confidence.

---

📊 Edge Benchmarking (PyTorch vs. ONNX vs. INT8)

Evaluated on standard CPU runtime across 50 iterations:

Runtime Engine	Precision	Model Size	Size Reduction	Mean Latency	P95 Latency	Throughput
PyTorch (Base)	FP32	107.8 MB	0.0%	73.7 ms	83.0 ms	13.6 FPS
ONNX Runtime	FP32	107.9 MB	0.0%	95.7 ms	145.6 ms	10.5 FPS
ONNX Runtime (Edge)	INT8	27.7 MB	74.3%	293.7 ms	403.9 ms	3.4 FPS

Numerical Parity Verification: Maximum absolute error between PyTorch FP32 and ONNX FP32 is $< 3.58 \times 10^{-6}$.

---

💻 Quickstart: Running Inference with ONNX Runtime

import numpy as np
from PIL import Image
import onnxruntime as ort

# Load Edge INT8 or FP32 model
session = ort.InferenceSession("unified_gated_plant_cnn_int8.onnx", providers=["CPUExecutionProvider"])

# Preprocess image
img = Image.open("leaf_sample.jpg").convert("RGB").resize((224, 224))
arr = np.array(img, dtype=np.float32) / 255.0
mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
std = np.array([0.229, 0.224, 0.225], dtype=np.float32)
tensor = np.transpose((arr - mean) / std, (2, 0, 1))[np.newaxis, :]

# Run single forward pass
disease_logits, crop_logits, saliency_mask = session.run(None, {"input": tensor})

# Results
pred_disease_idx = np.argmax(disease_logits, axis=-1)[0]
pred_crop_idx = np.argmax(crop_logits, axis=-1)[0]
print(f"Predicted Disease Index: {pred_disease_idx}, Crop Index: {pred_crop_idx}")

---

📜 Citation

@article{kumar2026plantcnn,
  title={Plant Leaf Disease Detection using Pure 2D CNNs with Differentiable Spatial Gating and Conformal Risk Control},
  author={Kumar, Shubham},
  year={2026}
}