AI Vision Quality Control

This YOLO-based computer vision model was delivered for an anonymized manufacturing organization during a professional assignment. At 5A Technologies, we build on the resulting computer-vision expertise. The model could locate objects in supplied images and video frames and visually mark them as conforming or non-conforming.

The delivery covered dataset validation and preparation, image annotation, local model training, evaluation and technical handover. The cameras, capture infrastructure and final production integration were outside this scope.

• Context: visual quality control in an anonymized manufacturing environment
• Project: professional machine-learning delivery
• Status: completed and technically handed over
• Role of 5A Technologies: carrying forward expertise in dataset validation, image annotation, local YOLO training, evaluation and handover
• Core principles: controlled data, human validation and iterative model improvement

Key design principle
Model quality depended not only on training, but also on manually validated images, consistent annotations and targeted feedback on incorrect or uncertain detections.

A computer vision model can learn only from what is visible and correctly labeled in the training data. The supplied image collection therefore first had to be made usable, consistent and representative before reliable model development was possible.

The main considerations were:

• identifying unusable, duplicate or unclear images;
• annotating conforming and non-conforming objects consistently;
• separating training, validation and test data;
• investigating false positives, false negatives and uncertain detections;
• translating feedback into improved labels and a new training cycle.

Validate the images

The supplied image collection was reviewed and cleaned first. Unusable, duplicate and unclear images were filtered before entering the model cycle.

Annotate the objects

Relevant objects received manually validated bounding boxes and labels. The dataset was then divided into separate training, validation and test data.

Train locally

A YOLO-based model was trained locally to locate objects in the supplied images and classify them as conforming or non-conforming.

Evaluate and refine detections

After each model candidate, incorrect, missed and uncertain detections were reviewed again. When corrections were needed, labels or training settings were adjusted and a new training and validation cycle followed.

1. Another party supplied the production images.
2. Unusable, duplicate and unclear images were filtered from the dataset.
3. Relevant objects received manually validated bounding boxes and labels.
4. The dataset was divided into separate training, validation and test data.
5. A YOLO-based model was trained locally to locate and classify objects.
6. Incorrect, uncertain and missed detections were investigated on validation and test images.
7. Corrected labels and user feedback informed a new training cycle where needed.
8. The selected model version was tested on images and video frames and technically handed over for further integration.

For each detection, the model output showed:

• where an object was located in the image through a bounding box;
• whether the object was classified as conforming or non-conforming;
• which confidence value was associated with the detection;
• which objects or regions required further human review.

This output supported model evaluation. No exact accuracy, precision, recall or mAP results are claimed as public project facts.

Confirmed core

• YOLO-based object detection: locating and classifying objects in supplied images;
• local model training: running training and validation cycles locally;
• human dataset validation: reviewing images and annotations before they entered the model cycle;
• iterative error analysis: using false positives, false negatives and uncertain detections to adjust labels and settings;
• images and video frames: checking the selected model version on still and moving visual input.

Historical technical reconstruction

The exact historical software versions and full toolchain are no longer confirmed. A plausible technical setup used Python, an Ultralytics YOLO or PyTorch environment, CVAT for annotation, OpenCV for image and video input, and a local NVIDIA GPU with CUDA for training.

These components describe a technical reconstruction, not a separately confirmed product or version contract. Precision, recall, mAP, false positives and false negatives are also relevant evaluation concepts, but no exact measurements are published without the original report.

Human review remained an active part of the model cycle. Incorrect labels, missed objects and uncertain detections could be corrected before the model was retrained.

The workflow was therefore not a one-off straight line. Feedback could return to the dataset and start a new training run, after which the model was checked again on separate validation and test images.

At 5A Technologies, we build on experience across:

• reviewing and cleaning the supplied image collection;
• validating and annotating relevant objects;
• structuring training, validation and test data;
• training the YOLO-based model locally;
• evaluating incorrect and uncertain detections;
• incorporating user feedback into new model cycles;
• testing the selected model version on images and video frames;
• technically handing over the model artifact and inference configuration.

The project delivered a technically transferable computer vision model that could locate objects in supplied images and video frames and visually mark them as conforming or non-conforming.

The combination of controlled annotations, local training and human feedback made it possible to evaluate and adjust model candidates deliberately before a final version was handed over to another party.

Because of the NDA and the bounded assignment, the public project description observes the following boundaries:

• the client, material, manufacturing process and defect types are not named;
• the original images are not shown;
• the cameras and capture infrastructure fall outside the delivered scope;
• the final production integration is not claimed as a 5A Technologies delivery;
• no exact dataset size or model version is published;
• accuracy, precision, recall, mAP, error rate, comparison with human inspection and operational production results are not presented as confirmed outcomes.

This project shows how a computer vision solution:

• treats dataset quality as a prerequisite for model development;
• embeds human validation in annotation and error analysis;
• deliberately separates training, validation and test data;
• uses feedback to refine labels, settings and new training cycles in a targeted way;
• hands over a model technically without adding unconfirmed production or performance claims.