Visual Recognition Pipeline for Industrial Parts & Fasteners

Objectives

What the project needed to achieve

• Evaluate zero-shot vs. fine-tuned accuracy across a 6,200 SKU mechanical fastener taxonomy
• Implement parameter-efficient LoRA tuning targeting adapter sizes under 100 MB
• Evaluate training duration and single-workstation compute requirements for parameter-efficient adaptation
• Evaluate top-1 visual classification accuracy improvements over zero-shot baselines

Challenge

Engineering constraint

Identifying mechanical fasteners visually is difficult due to untextured metallic finishes, subtle thread pitch differences, and head geometry variations (hex, torx, socket). Training custom classifiers from scratch requires massive datasets, while full fine-tuning of large foundation models creates gigabyte-sized weights that are expensive to deploy. This study evaluated whether lightweight LoRA adapters can achieve precision recognition at low training cost.

Approach

How YantriX approached the work

1. 01

   Assembled a structured dataset of mechanical fasteners under varied industrial illumination.

2. 02

   Evaluated zero-shot OpenCLIP baseline across fastener categories to establish a reference floor.

3. 03

   Injected rank-16 LoRA adapters into vision transformer projection layers, freezing backbone weights.

4. 04

   Executed fine-tuning using PyTorch and HuggingFace PEFT libraries on a single workstation GPU.

5. 05

   Evaluated inference latency and adapter swapping performance in an automated test harness.