GENERATION OF SYNTHETIC IMAGES FOR TRAINING OF DEFECT DETECTION MODELS

Smart overview of the Invention

The patent application describes a system for generating synthetic images to train defect detection models, specifically for pallet defects. This system uses a small set of real images with defects, along with associated masks and textual descriptions, to fine-tune a latent diffusion model. The model generates synthetic images by inpainting defects into masked regions of input images. These synthetic images are then filtered to remove outliers, ensuring they match real defect distributions, and are used alongside real images to train an object detection model.

Background

Automatic defect detection is crucial in various industries, including manufacturing and shipping. In the global pallet inspection industry, it helps reduce human error and improve efficiency. However, training models for defect detection requires a large dataset of images with defects, which is challenging to obtain due to the prevalence of undamaged pallets. Conventional data augmentation techniques are limited as they do not accurately model defects, making the generation of synthetic images a valuable solution.

Technical Approach

The system employs a diffusion model to learn defect distributions and generate synthetic images with realistic defect regions. This generative AI approach allows for the creation of synthetic pallet defect images, enabling the training of object detection models with sufficient data to achieve desired accuracy. The model can simulate new defects on undamaged pallets or modify existing defects, enhancing the diversity and quantity of training data.

Implementation Details

The technique involves fine-tuning a latent diffusion model using a small set of defect images and their masks. This refined model generates synthetic images by inpainting defects, which are then filtered to maintain distribution accuracy. The resulting dataset, combining filtered synthetic images with real defect images, is used to train an object detection model. This approach reduces the need for extensive real-world data and improves defect detection accuracy.

Advantages

The system offers rapid generation of training data, significantly reducing the time required compared to waiting for real defects. It enhances detection accuracy and efficiency, needing only a few annotated images for model fine-tuning. The approach allows for targeted defect injection in realistic locations, improving the effectiveness of defect detection applications across various industries. The system is implemented using programmable circuitry, allowing for flexible deployment in environments like warehouses.