[Paper] Adaptive Hybrid Optimizer based Framework for Lumpy Skin Disease Identification

Source: arXiv - 2601.01807v1

Overview

The paper introduces LUMPNet, a hybrid deep‑learning framework that combines object detection (YOLOv11) and image classification (EfficientNet) with a newly designed adaptive hybrid optimizer to spot Lumpy Skin Disease (LSD) lesions on cattle. By automating early detection from photographs, the approach promises faster, more reliable disease monitoring for farmers and veterinary services.

Key Contributions

• Hybrid Architecture: Integrates YOLOv11 for nodule localization with an EfficientNet‑based classifier for disease vs. healthy decision.
• Adaptive Hybrid Optimizer: A custom optimizer that blends the strengths of AdamW and SGD‑style momentum to stabilize and speed up training of both detection and classification heads.
• Compound Scaling of EfficientNet: Applies EfficientNet‑B0/B1 scaling rules to balance model size, accuracy, and inference speed on edge devices.
• Empirical Validation: Achieves 99 % training accuracy and 98 % validation accuracy on a public LSD image dataset, outperforming prior CNN‑only baselines.
• Case‑Study Comparison: Demonstrates that the full LUMPNet pipeline exceeds a stand‑alone EfficientNet‑B0 model trained with AdamW, confirming the benefit of the hybrid design.

Methodology

1. Data Preparation

   • Publicly released cattle skin image set (healthy and LSD‑affected).
   • Images are resized to 640 × 640 px for YOLOv11 and 224 × 224 px for EfficientNet.
   • Standard augmentations (random flip, rotation, color jitter) are applied to improve robustness.

2. Detection Stage (YOLOv11)

   • YOLOv11 scans the whole image and outputs bounding boxes around suspected nodules.
   • Confidence thresholds are tuned to keep high‑recall detections while limiting false positives.

3. Classification Stage (EfficientNet)

   • Cropped patches from YOLO’s bounding boxes are fed into an EfficientNet backbone.
   • The network uses compound scaling (depth, width, resolution) to keep the model lightweight for on‑farm devices.

4. Adaptive Hybrid Optimizer

   • Starts with AdamW’s adaptive learning‑rate per‑parameter for rapid early convergence.
   • Switches to an SGD‑with‑momentum regime after a preset epoch or when validation loss plateaus, reducing over‑fitting and improving generalization.
   • Learning‑rate schedules (cosine decay) are applied in both phases.

5. Training & Evaluation

   • Joint training of detection and classification heads using a multi‑task loss (YOLO objectness + EfficientNet cross‑entropy).
   • Metrics: mean Average Precision (mAP) for detection, accuracy & F1‑score for classification.

Results & Findings

• The hybrid optimizer reduced training epochs from 120 to 85 while maintaining higher validation scores.
• YOLOv11’s precise nodule localization trimmed the amount of data the classifier needed to process, cutting inference latency.
• Ablation studies confirmed that removing either the detection stage or the optimizer switch caused a 3–5 % drop in overall accuracy.

Practical Implications

• Field‑Deployable Diagnostics: The compact model can run on low‑cost edge devices (Raspberry Pi, Jetson Nano), enabling veterinarians to scan cattle on‑site with a smartphone camera.
• Early Outbreak Containment: Real‑time alerts can be integrated into farm management software, triggering quarantine or treatment protocols before the disease spreads.
• Scalable Monitoring: Cloud‑backed pipelines can ingest images from multiple farms, automatically flagging high‑risk herds and feeding data into epidemiological dashboards.
• Transferable Framework: The detection‑plus‑classification pattern, coupled with the adaptive optimizer, can be repurposed for other livestock diseases (e.g., foot‑and‑mouth, bovine tuberculosis) that manifest as localized lesions.

Limitations & Future Work

• Dataset Diversity: The public dataset contains limited breeds and lighting conditions; broader field data may expose robustness gaps.
• Hardware Constraints: While inference is fast on modest CPUs, extreme low‑power IoT nodes may still struggle without further model pruning or quantization.
• Optimizer Generalization: The hybrid optimizer’s switching criteria were manually set; an automated schedule (e.g., based on loss curvature) could improve adaptability across tasks.
• Multi‑Modal Extension: Incorporating non‑visual data (temperature, movement patterns) could boost detection confidence, a direction the authors plan to explore.

Authors

• Ubaidullah
• Muhammad Abid Hussain
• Mohsin Raza Jafri
• Rozi Khan
• Moid Sandhu
• Abd Ullah Khan
• Hyundong Shin

Paper Information

• arXiv ID: 2601.01807v1
• Categories: cs.CV, cs.AI
• Published: January 5, 2026
• PDF: Download PDF