<p>Advancements in vision inspection technology have improved quality management in smart manufacturing. However, most existing approaches focus on classifying products as defective or nondefective, without quantifying the severity of defects or explaining model decisions. This study proposes an approach that combines deep learning with explainable AI (XAI) to quantify the degree of quality risk within the region of interest (ROI). Extending the role of XAI beyond reliability analysis, the proposed approach converts pixel-level attributions into object-level severity reasoning and expresses the result as a continuous quality risk score (QRS) between 0 and 1. Furthermore, by requiring only binary-labeled data (i.e., defective vs. nondefective), the approach eliminates the need for fine-grained annotations. The proposed approach was validated using the proprietary semiconductor and Sensum solid oral dosage forms datasets. Across all tested models and datasets, defective products exhibited significantly higher QRS than nondefective products, and all differences were statistically significant at the 0.01 level. In addition, the method achieved a broader score dispersion than previous confidence-based methods and consistently showed higher coverage across all shrinking factor values, with an average of up to 34.16 percentage points higher coverage across the evaluated datasets. This wider distribution reflects the enhanced ability to distinguish subtle quality differences and supports its use as a practical basis for grading and selectively allocating products.</p>