Deep LearningCompletedJan 2026

YOLOv11 Architecture Optimization

S1 thesis modifying YOLOv11 with Coordinate Attention on the neck module — improving small PPE object detection for industrial settings.

Source code

Role: Researcher & Implementer
Timeline: Sep 2025 – Jan 2026
Team: Solo
Client: Universitas Brawijaya

The problem

Object detectors routinely miss small objects — a distant or partly hidden helmet — which is exactly what industrial safety monitoring cannot afford. My S1 thesis asked whether attention in YOLOv11's neck could close that gap.

Constraints

Improve small-object detection without the compute cost of a much larger model
Evaluate on industrial PPE data, against the unmodified YOLOv11 baseline
One researcher, one thesis cycle

Architecture

Step 1, Data
Industrial PPE dataset
Training and evaluation data
Step 2, Compute
YOLOv11 backbone
Unchanged feature extractor
Step 3, Compute
Neck + Coordinate Attention
Positional and channel attention
Step 4, Compute
Detection head
Unchanged YOLOv11 head
Step 5, Interface
Benchmark vs baseline
PyTorch runs, Pandas + Matplotlib analysis

YOLOv11's backbone and detection head are kept as they are; a Coordinate Attention module is inserted into the neck, and the modified and stock models are trained and benchmarked on the same industrial PPE dataset.

Engineering decisions

Coordinate Attention in the neck
Context
Channel attention alone says what to focus on but not where — and small objects need positional detail to survive feature aggregation.
Decision
Inserted a Coordinate Attention module into the neck so the network encodes positional information alongside channel attention.
Trade-off
One more module in the architecture, kept cheap because Coordinate Attention factorises spatial attention into two 1D encodings.
A targeted change over scaling up
Context
Scaling the whole model would add parameters everywhere without addressing why small objects were being missed.
Decision
Studied the baseline's failure modes first, then changed only the part of the network responsible.

Outcomes

Architecture Change

Task

Domain

Measurable gains on small PPE object detection over the stock YOLOv11 baseline
Completed S1 thesis at Universitas Brawijaya, with a final GPA of 3.58 / 4.00

Reflections

Lessons learned

The clearest takeaway was that targeted architectural changes, grounded in the actual failure modes of the baseline, beat blindly scaling the model. Understanding why small objects were being missed in the first place is what told me where attention would genuinely help — and where it would just add parameters.