Date of Award

6-2026

Degree Name

MS in Electrical Engineering

Department/Program

Electrical Engineering

College

College of Engineering

Advisor

Lynne Slivovsky

Advisor Department

Electrical Engineering

Advisor College

College of Engineering

Abstract

A custom convolutional neural network (CNN) for automated cosmetic condition grading of used electronics achieves a macro-averaged F1 score of 0.419 ± 0.075 and accuracy of 0.490 ± 0.055, outperforming a zero-shot baseline using Google's Gemini vision-language model (VLM) (macro-F1 of 0.370 ± 0.135, accuracy of 0.420 ± 0.152) with roughly half the fold-to-fold variance. The IT Asset Disposition (ITAD) industry processes millions of retired electronic devices annually, assigning cosmetic condition grades that assist in determining whether devices are resold, refurbished, or recycled. This grading is performed manually by trained technicians, a process that is inherently subjective and inconsistent across operators. Despite the economic significance of accurate grading, no prior work has applied computer vision to automate whole-device cosmetic condition assessment. This thesis addresses that gap with a late-fusion CNN using a MobileNetV2 backbone that processes three standardized views per device (bottom, top lid, and open) before concatenating features for classification. A dataset of 100 Chromebooks (300 images) was collected at a partner ITAD facility under real warehouse conditions, with each device graded into one of three categories (B, C, or D) by trained technicians. Both approaches are assessed via five-fold stratified cross-validation with macro-averaged F1 score as the primary metric. An ablation study identifies data augmentation as the most critical design factor, with its removal causing a 32% relative decline in macro-F1, while ImageNet-pretrained transfer learning yields the best accuracy-stability tradeoff. These results represent the first application of deep learning to automated cosmetic grading in ITAD and suggest that confidence-based routing, combined with expanded training data, offers a viable path toward practical deployment.

Available for download on Friday, June 01, 2029

Share

COinS