Hitachi unveils selective re-encoding that halves ViT image AI processing time

Hitachi's announcement on August 25, 2026, centers on a technical bottleneck that has long affected image analysis with AI: Vision Transformer (ViT) models generate tokens for the entire image, which increases processing time and computational load. The new selective re-encoding technology addresses this by identifying which regions of the image contribute to fixed regions in the model's internal representation, and re-encoding only the tokens corresponding to those regions. This approach allows the system to maintain accuracy while reducing the number of tokens it must process.

The company verified the effect through trial calculations assuming image analysis, finding that the average processing time was halved while accuracy was maintained. This is particularly relevant for OT (operational technology) fields that use large amounts of image and video data, such as defect detection in manufacturing or abnormality monitoring from surveillance camera footage, where reducing GPU and other computational resource requirements could enable more efficient AI deployment.

Hitachi says it will now evaluate the technology's value in actual use cases, starting with image and video recognition AI, and expand application to AI and social infrastructure, linked to its AI-driven strategy plan 'Lumada 3.0'. The outcome hinges on whether real-world deployments can replicate the trial results and whether the approach integrates smoothly with existing infrastructure. For teams managing image-heavy AI workloads, the promise of halved processing time could ease resource constraints, though the full impact will depend on adoption and scaling beyond the initial verification.