↓Skip to main content

Chickpea Fusarium Wilt Detector

Table of Contents

MSc Thesis Research | Batangas State University | 2024–2026
Published Paper: A Systematic Ablation of CBAM Depth in SqueezeNet for Optimising Cross-Domain Chickpea Disease Detection, Procedia Computer Science, Vol. 289, pp. 1259–1266, 2026 (ICCSCI 2026). DOI: 10.1016/j.procs.2026.08.346

What It Does

A production-ready Android application that detects Chickpea Fusarium Wilt from smartphone photos — fully offline, with instant inference results and bilingual disease management advisory (English + Myanmar Unicode). Designed specifically for smallholder farmers in rural Myanmar with no internet connectivity.

Key Results

  • 82.09% cross-domain accuracy on an unseen real-world Myanmar field test set, outperforming MobileNetV2, EfficientNet-B0, and ShuffleNet.
  • 0.870 MB model size after static INT8 quantization (a 3.31× compression factor from the 2.88 MB FP32 baseline) with zero accuracy degradation.
  • Validated across 115 real physical Android devices via Google AI Edge Portal (private preview).
  • Sub-75ms worst-case inference (73.7 ms) on entry-level CPU-only hardware, compatible with 2 GB RAM devices and Android 9.0+.

Technical Contributions

  • Novel Architecture: Designed a hybrid CNN combining SqueezeNet v1.1 with Convolutional Block Attention Modules (CBAM), conducting systematic ablation across network depths to identify optimal attention placement.
  • Rigorous Quantization: Engineered an INT8 static quantization calibration pipeline tailored for edge tensor runtimes (LiteRT / TensorFlow Lite).
  • Primary Field Dataset: Personally collected and curated field imagery from chickpea-growing farming communities across Myanmar, annotated in collaboration with professional agronomists and validated using Cohen’s kappa coefficient.
  • Bilingual Mobile Application: Built a modern, native Android client using Kotlin, Jetpack Compose, Material Design 3, and LiteRT.

Why It Matters

Most state-of-the-art vision models require high-end GPUs or round-trip cloud API connectivity. In rural Southeast Asia, smallholder farmers often lack reliable mobile data or high-end smartphones. This work delivers research-grade computer vision diagnostics to budget devices under challenging field conditions.

Tech Stack: PyTorch · LiteRT / TFLite · Kotlin · Jetpack Compose · Android Studio · Roboflow · Kaggle GPU · Google AI Edge Portal