When good parts are incorrectly rejected, scrap costs rise, rework capacity is tied up, and delivery deadlines are put at risk. With robust yet flexible inspection logic, pseudo-scrap can be reduced sustainably.
In this article, you’ll learn what causes pseudo scrap, what it costs, and how the False Rejection Rate (FRR) is measured and reduced for the long term. We also take a closer look at the role of AI in quality inspection.
In quality inspection, there are two types of errors: parts that are incorrectly rejected as bad, and parts that are incorrectly accepted as good. These are referred to as pseudo scrap (false rejects) and escapes (false accepts).
When products are rejected during or after the manufacturing process by quality inspection, this is considered scrap. If, however, parts that meet the quality standards are mistakenly rejected, this is pseudo scrap (equivalent to FRR—False Rejection Rate).
Escape is the opposite: products pass inspection even though they have defects (equivalent to FAR—False Acceptance Rate). Both misclassifications create costs and inefficiencies in production.
Incorrectly rejected parts (i.e., a high False Rejection Rate) create direct and indirect costs across the entire value chain.
Pseudo scrap impacts profitability through avoidable material, energy, and process costs. Resources have already been invested in manufacturing, labor, and machine runtime—yet good product is reworked or even discarded in the end. This reduces usable output and ties up capacity that could create value elsewhere.
Economic losses: Every incorrectly rejected item represents a direct loss of material, labor time, and margin.
Productivity loss: Incorrect sorting reduces the true output of the production line because some flawless product never becomes saleable inventory.
Inefficient use of resources: Personnel and equipment dedicated to rework or disposal of pseudo scrap are used inefficiently.
False rejects directly affect two central manufacturing KPIs: Yield and OEE. Yield is the share of produced parts that actually go to sale as good parts, while OEE (Overall Equipment Effectiveness) describes how effectively equipment is utilized across availability, performance, and quality.
Incorrectly rejected good parts lower the quality rate and therefore reduce yield and OEE. Energy and upstream value-add already invested no longer create value, and usable output declines.
In quality inspection, two objectives collide: letting no defective parts through and not stopping good parts unnecessarily. The stricter the inspection, the fewer real defects slip through—but the number of incorrectly rejected good parts (pseudo scrap) increases. More permissive settings allow more good parts to pass, but the risk that defective parts remain undetected (escapes) also rises. The economically optimal point is where total costs from misclassification, rework, and claims are minimized.
The FRR–FAR trade-off exists as long as the features of good and bad parts overlap. It can only be resolved when inspection logic can clearly separate good from bad so that a clear threshold exists without conflicting goals. Data-driven algorithms such as AI can achieve this by modeling real-world complexity with high accuracy.
A perfectly clear-cut decision between good and bad parts is rarely possible in practice. Pseudo scrap can have many causes. The following provides a structured overview of the most important factors:
Reflections, gloss, changing surfaces, and stray ambient light alter contrast and create apparent features that can make good parts look defective.
Dust, fibers, oil residues, or “trash material” on parts, carriers, or glass plates can mask features or trigger false alarms.
Subjective or ambiguous evaluation criteria and complex defect patterns overwhelm human inspectors and rule-based systems, increasing FRR.
Incorrect parameter settings, faulty or missing recalibration, and changed process conditions shift scores toward rejection.
Position variation, vibration, and timing errors in triggering or exposure reduce capture repeatability and increase misclassifications.
Companies reduce the false reject rate in visual inspection using three levers: improving image quality, optimizing data and algorithms, and using early warnings to detect deviations at an early stage.
The goal of the inspection setup is consistent, repeatable image capture where features are clearly visible—so lighting and optics are the main focus. Diffuse, coaxial, or dome lighting can reduce reflections and stabilize contrast.
Polarizing filters and shielding minimize stray light and improve repeatability. Constant cycle times and stable trigger and shutter signals reduce variation in captured images. Cleanliness also directly affects misclassification: blowing off or extracting contamination before image capture can be helpful.
Rule-based programs reach their limits with gloss, texture changes, and product variants. Continuously changing conditions and complex defect patterns make classification with rigid rules difficult.
That’s why modern industrial vision increasingly relies on AI. It learns robust decision boundaries from real data to separate good and bad parts. A consistent ground truth with golden samples is essential—i.e., referenced sample parts that are clearly released as “good” or “bad,” especially defining the borderline cases. This determines what the system will later recognize as a good part or a defect.
AI can also be adapted easily to new production conditions: instead of reprogramming by specialists, the training dataset is simply updated—for example by adding samples from the latest production batch.
Even after deployment, regular verification of model performance remains critical. Continuous monitoring of inspection accuracy makes production issues visible early, while notifications when scrap rates rise enable timely corrective action.
Pseudo scrap (False Rejection Rate) increases costs and slows production efficiency. Because the causes are diverse, the most effective approach is a combination of three levers:
First, design image quality so all relevant features and defects are reliably visible. Second, establish a consistent data foundation (ground truth), ideally including borderline samples. Third, continuously monitor inspection algorithms in production and adjust them as needed.
AI-based methods improve separation performance compared to rule-based systems. They can also be retrained without deep expert knowledge and are well suited for retrofit scenarios.
Discover how easily you can solve a wide range of inspection tasks with Maddox AI.
Explore Maddox AI software
We are ISO 27001 certified.
