Synthetic Pallet Defect Images Using Diffusion Inpainting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Obtaining sufficient training data for defect detection models is challenging due to the unbalanced nature of datasets, with many more undamaged pallets than damaged pallets, and existing data augmentation techniques fail to effectively model actual object defects.
Innovation Solution
Generate synthetic images using a diffusion model to learn defect distributions and inpaint defects on undamaged regions, refining a latent diffusion model with a small number of training images, and filter outliers to create a training dataset for object detection models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional image processing techniques are used for data augmentation, then the process is simple and fast, but the training data does not effectively model actual object defects
Solution Approach 1:
The patent uses diffusion models to generate synthetic images that copy and simulate actual defect patterns from real pallet images. The model learns from real defect data and creates realistic synthetic defects, allowing the system to train on abundant simulated defect images without needing to physically create or photograph every possible defect type.
Solution Approach 2:
The diffusion model acts as an intermediary between real defect data and training data generation. It takes real pallet images with defects, learns the defect distribution patterns, and generates synthetic images that bridge the gap between limited real data and the need for extensive training data, thereby improving detection accuracy without requiring direct manipulation of physical pallets.
2Quantity of substance
If real pallet images are collected for training, then the data represents actual defects, but it takes an unacceptable amount of time and is costly to annotate
Solution Approach 1:
The system performs preliminary action by training a diffusion model on a small set of real defect images first. Once the model learns the defect distribution patterns from this initial data, it can rapidly generate large quantities of synthetic training images without requiring continuous collection and annotation of real defect data, thus reducing time loss while maintaining data quality.
Solution Approach 2:
The diffusion model creates synthetic copies of real defect patterns by learning from limited real images. These synthetic copies can be generated in large quantities rapidly, providing abundant training data without the time-consuming process of collecting and manually annotating additional real pallet images with defects.
3Productivity
If diffusion models are used to generate synthetic images, then training data quantity increases rapidly, but the model requires computational resources and tuning
Solution Approach 1:
The patent applies parameter changes by tuning the diffusion model's hyperparameters and training configurations to optimize the balance between generation speed and computational consumption. By adjusting parameters such as the number of diffusion steps, image resolution, and model architecture, the system achieves rapid synthetic image generation while managing computational resource requirements.
Data Source
AI summary
Systems, apparatus, articles of manufacture, and methods to generate synthetic images for training of defect detection models are disclosed. An example system disclosed herein produces synthetic images of pallet defects to train object detection models. In some examples, a small set of real images containing defects, associated masks and textual descriptions is used to fine tune a latent diffusion model. The fine-tuned model accepts a masked input image, a mask that defines the region to be altered, and a defect description, and generates a synthetic image with the defect inpainted into the masked region. In some examples, generated synthetic images are filtered to remove outliers that do not match the real defect distribution. The filtered synthetic dataset, together with a limited set of real images, is then used to train a downstream object detection model capable of identifying pallet damage in captured images.