Microplastic Identification using AI-Driven Image Segmentation and GAN-Generated Ecological Context.

§ 01 — TL;DRWhat we did.

A segmentation model that identifies microplastics in real-world imagery, trained with synthetic ecological context generated by a GAN — so it generalizes across backgrounds it has never seen.

Microplastic detection in the wild is hard because the foreground (a fragment of plastic) looks deceptively similar to the background (sand, kelp, sediment, foam). Models trained on tightly-curated lab images learn a shortcut: they classify the background, not the object. We use a GAN to manufacture diverse, realistic backgrounds and re-paste real microplastic masks onto them, expanding the training set in a way that stresses the model along the axis that actually matters at deployment.

§ 02 — The problemBackground bias.

Two things break microplastic detection:

1. Scarce, narrow data. Annotated microplastic datasets exist, but they tend to be captured under controlled lighting, on uniform substrates. Field samples look nothing like that.
2. Foreground / background entanglement. Plastic shards on sand have low contrast. A CNN learns "this kind of speckled brown texture means plastic" because that's what the training pairs taught it — and then misses the same shard on a wet rock.

§ 03 — ApproachAugment with the world, not with noise.

Rather than augmenting with rotations and color jitter — which leave the spurious foreground/background pairing intact — we augment with plausible ecological context. A generative model produces unseen-but-realistic backgrounds (sand textures, kelp tangles, wet pebbles, foam). We composite the real microplastic masks onto these new backgrounds and add the composites to training.

The intuition is simple: if the only thing that stays constant between two training images is the object itself, the model has to actually learn the object.

Fig. 01 Segmentation outputs on field samples. The model produces clean masks across substrate types (sand, sediment, organic debris) it never saw in the original lab dataset.

§ 04 — PipelineFour stages.

The composite step is where most of the leverage lives. Naive copy-paste creates an obvious edge artifact the model can exploit; we use blending and lighting consistency so the only reliable signal is the object's shape and material.

§ 05 — ResultsGeneralization, not accuracy theater.

The headline isn't a higher in-distribution score — it's that the same model holds up on substrates and lighting conditions absent from the original training set. We test on held-out field photography and report cross-substrate IoU.

Notably, the in-distribution metric barely moves. That's the point. Augmentation that chases benchmark score often improves it at the cost of robustness; the GAN-context augmentation here trades nothing in-distribution and buys a real chunk of generalization.

§ 06 — TakeawaysWhat I'd want a reader to leave with.

• Augmentation is a model of the world.

  "Add noise" implies failures are random. They aren't — they're structured by what the model expects to see. Augment along the structure.
• Generative models can be training infrastructure.

  A GAN here isn't the product; it's a controllable source of diversity for the supervised model that is the product.
• Robustness needs its own metric.

  Cross-substrate IoU surfaces what an aggregate score hides. The gap between "test score" and "deployment score" is the real number.
• Domain context beats domain randomization.

  Random textures don't help. Plausible textures do — because the model's failure surface is shaped by plausibility, not by entropy.

§ 07 — CiteFor the bibliography.

If this work was useful to you:

@article{dils2024microplastic,
  title   = {Microplastic Identification Using AI-Driven Image Segmentation
             and GAN-Generated Ecological Context},
  author  = {Dils, Alex and Raymond, David and Sad{\'e}e, Christoph and others},
  journal = {arXiv preprint arXiv:2410.19604},
  year    = {2024}
}

← Back to Research · Index