ICAO-Compliant Facial Images
- ICAO-compliant facial images are standardized portrait photographs designed to meet ICAO Doc 9303 and ISO/IEC requirements for biometric verification in travel documents.
- They enforce strict constraints on pose, expression, illumination, and background to ensure machine readability and seamless interoperability in identity workflows.
- Automated compliance verification leverages multi-task deep models and quality assessment tests to enhance image security, resilience against manipulation, and operational efficiency.
ICAO-compliant facial images are standardized portrait photographs intended for biometric use in Machine Readable Travel Documents (MRTDs), electronic MRTDs (eMRTDs), and closely related identity-verification workflows. Their defining property is conformance to the portrait-quality and interchange constraints specified by ICAO Doc 9303 together with ISO/IEC face image standards, notably ISO/IEC 19794-5 and ISO/IEC 39794-5. In practice, this means that pose, expression, illumination, background, occlusion, spatial sampling, and image coding are constrained so that the stored reference image remains interoperable, machine-readable, and suitable for automated face verification at border control, document issuance, and adjacent digital identity systems (Domenico et al., 2024, Merkle et al., 2022).
1. Normative basis and scope
ICAO-compliant facial imagery is rooted in a layered standards stack. ICAO Doc 9303 defines the biometric role of the face in MRTDs and the associated logical data structures, while ISO/IEC 19794-5 and its successor ISO/IEC 39794-5 define face image data formats and capture constraints for interoperability in biometric systems. ISO/IEC 29794-5, and the associated OFIQ effort, operationalize quality assessment by translating these constraints into measurable image-quality components (Chivata et al., 26 Aug 2025, Merkle et al., 2022).
| Document or framework | Primary function | Typical constraints or outputs |
|---|---|---|
| ICAO Doc 9303 | MRTD deployment and interoperability | Face as primary biometric |
| ISO/IEC 19794-5 / 39794-5 | Face image data and capture rules | Pose, expression, background, illumination |
| ISO/IEC 29794-5 / OFIQ | Quantitative face image quality assessment | Metrics for blur, lighting, occlusion, positioning |
The portrait-quality parameters summarized in the deep data hiding survey are explicit: inter-eye distance should be at least 90 pixels, with 120 pixels recommended; the background should be uniform and light-colored with no patterns; lighting should be uniform with no strong shadows; pose should be full frontal with the head vertically aligned; expression should be neutral with mouth closed; printed-image saturation should keep non-background pixels at values 0 or 255 below 0.1%; scanned portraits should be at least 300 dpi; and JPEG or JPEG 2000 are the document-relevant compression formats (Chivata et al., 26 Aug 2025).
These rules are not merely aesthetic. They define the admissible operating region for interoperable biometric recognition. A plausible implication is that ICAO compliance should be understood as a constrained acquisition-and-representation regime rather than as a single binary property of the final bitmap.
2. Capture geometry and portrait-quality constraints
The operative constraints fall into three broad groups: geometric, photographic, and pose/facial-attribute constraints. In ONOT, these are implemented as 21 SDK tests covering uniqueness and inclusion of the face, eye distance, horizontal and vertical facial position, head-width and head-height ratios, focus, saturation, illumination, glare on glasses, gaze direction, mouth expression, shoulder position, eye visibility, occlusions, and related portrait attributes. An image is treated as compliant only if every test score satisfies
Several studies instantiate these requirements through explicit normalization. In the retouching-compression study, manually selected reference images were required to be good-quality frontal faces with mostly neutral expression and to meet ICAO specifications; in particular, “an inter-eye-distance of at least 90 pixels in the facial image has to be fulfilled.” After landmark detection with dlib, images were aligned so that the eyes occupied fixed positions, the intra-eye distance was 90 pixels, and the resulting image size was 360×480 pixels (Rathgeb et al., 2021).
A more granular decomposition appears in FaceQvec and DFIC. FaceQvec operationalizes 25 tests linked to ISO/IEC 19794-5, including blur, eyes direction, ink marks, odd skin colour, illumination, contrast, pixelation, hair over face, eyes open/closed, heterogeneous background, pose estimation, light reflections on skin, red eyes, shadows in background, shadows over face, sunglasses, reflections on glasses, wide frames, frames covering the eyes, hats, veils, mouth open/closed, other faces, white-noise estimation, and facial expression (Hernandez-Ortega et al., 2021). DFIC extends the requirement set to 26 conditions, explicitly enumerating eyes closed, non-neutral expression, mouth open, rotated shoulders, roll/pitch/yaw, looking away, hair across eyes, head coverings, veil over face, other faces or toys/objects, dark tinted lenses, frame coverings of the eyes, flash reflections, heavy frames, shadows behind head, shadows across face, flash reflection on skin, unnatural skin tone, red eyes, too dark/light, blur, varied background, pixelation, washed out appearance, ink marked/creased artifacts, and posterization (Gonçalves et al., 11 Feb 2026).
This decomposition is important because ICAO compliance is often locally violated. The image may satisfy global pose and resolution conditions while failing on eye visibility, flash reflection, or background uniformity. That local structure strongly influences how modern compliance-verification systems are designed.
3. Operational role in travel documents and border control
ICAO-compliant facial images function as the reference biometric in passports and related credentials. At automated border control (ABC) gates, the eMRTD chip image is compared against a live capture; the stored image is therefore expected to be frontal, neutral, and standardized, while the live probe may be only semi-controlled. The eyebrow-based morphing paper describes the standard ABC workflow as: read the chip, extract the ICAO-compliant facial reference image, capture a live image, and perform face recognition between the two (Zafar et al., 2023).
Compression is intrinsic to this operational setting. The retouching study explicitly motivates its experiments by noting that image compression is applied to facial images in electronic travel documents, and it uses JPEG and JPEG 2000 settings chosen to match ICAO recommendations for ePassport storage (Rathgeb et al., 2021). The recent 1024-byte study extends the same logic to temporary travel documents, arguing that 2D barcodes could provide machine readability when RFID chips are absent, but then the facial image must be compressed to at most 1024 bytes while the reference sample must still comply with ICAO specifications (Andreas et al., 29 Jun 2026).
The operational distinction between reference and probe images is technically consequential. In the retouching-compression study, reference images are ICAO-compliant and probe images are unconstrained. In differential settings, the comparison pair is therefore asymmetric by design: one side is a standardized passport-style portrait and the other is a live or semi-controlled capture (Rathgeb et al., 2021). A plausible implication is that ICAO compliance is less a property of all images in a recognition transaction than of the reference anchor around which the transaction is organized.
The same asymmetry appears in quality assessment for surveillance and screening. The lightweight Random-Forest-based framework for real-time screening uses a high-quality reference image together with lower-quality CCTV probes, and demonstrates that pre-filtering on quality can substantially improve downstream ArcFace verification (Ibrahim et al., 21 Jul 2025). Although this framework targets surveillance rather than passports, its use of frontal, high-quality reference imagery parallels the role of the ICAO portrait in document-centric verification.
4. Automated compliance verification and quality assessment
Automated ICAO verification has evolved from rule-based SDKs and per-component test suites toward multi-task deep models and hybrid standards-driven FIQA systems. FaceQvec is representative of the component-wise tradition. It outputs a 25-dimensional binary quality vector,
where each component records conformance to a specific ISO/IEC 19794-5 requirement, allowing a system not only to reject an image but also to diagnose which factors failed (Hernandez-Ortega et al., 2021).
The DFIC work moves further toward fully automatic multi-label validation. DFIC contains around 58,000 annotated images and 2706 videos of more than 1000 subjects, including both compliant portraits and a broad range of non-compliant conditions, with a nearly uniformly distributed partition for demographic balance. On that basis, the authors fine-tune a spatial-attention-based model with segmentation-guided region masking and requirement-wise weighted BCE losses. On the DFIC test set, this model achieves an average EER of 0.020 across 26 requirements; on TONO+ONOT it achieves 0.036; and its reported Bias Index is 0.047, reduced to 0.038 under balanced training (Gonçalves et al., 11 Feb 2026).
A lighter-weight engineering approach appears in the CCTV screening framework, where normalized facial landmarks and a Random Forest are used to score primarily pose and face resolution. On a real-world dataset of over 600 subjects from Dubai Police CCTV, the quality model reaches 96.67% accuracy. When inserted before ArcFace, mean cosine similarity for genuine pairs improves from 0.66 to 0.76, and FRR drops from 13.19% to 0.04% (Ibrahim et al., 21 Jul 2025). This suggests that even partial ICAO-like quality gating—centered on pose and size rather than full portrait compliance—can materially stabilize verification.
The fairness question is non-trivial. In a study of all ISO/IEC 29794-5 measures across age, gender, and skin tone, most measures showed no clear bias toward any specific demographic group, but dynamic range and luminance mean exhibited considerable variation across skin tone groups (Kabbani et al., 14 Jan 2025). A dataset-specific result in MORPH is more cautionary: using IFace SDK compliance checks, just over 48% of the African-American cohort and just over 57% of the Caucasian cohort were ICAO-compliant, and the brightness element differed significantly between the cohorts (Krishnapriya et al., 2019). The combined evidence suggests that automated compliance systems should separate phenotype-sensitive photometric effects from genuine acquisition failures.
5. Compression, storage, and recognition trade-offs
Compression is inseparable from ICAO-compliant imagery because the reference portrait is routinely stored under explicit storage constraints. In the ePassport-oriented retouching study, ICAO guidance is summarized as a minimum practical image size of approximately 12 kB and an operational optimum in the 15–20 kB range; the experiments therefore compress ICAO-compliant reference images to 15 kB with JPEG 2000 and 20 kB with JPEG (Rathgeb et al., 2021).
At these ICAO-level settings, compression interacts differently with different feature families. For BSIF texture descriptors, JPEG 2000 at 15 kB can severely degrade retouching detection; the paper’s illustrative AirBrush cross-database example reports D-EER values of 1.11% for the original image, 1.11% after JPEG, and 18.07% after JPEG 2000 in a single-image setting. By contrast, ArcFace-based deep face representations remain almost unchanged: averaged single-image D-EER is about 20.71% for original images, 20.72% for JPEG, and 20.63% for JPEG 2000; averaged differential D-EER is about 12.58%, 12.68%, and 12.66%, respectively (Rathgeb et al., 2021).
The same study also reports that the best overall performance is obtained in the differential scenario using ArcFace embeddings, around 12.5% D-EER across apps and databases, closely matching the operational case in which a compressed ICAO-compliant passport image is compared with a live probe (Rathgeb et al., 2021). This suggests that compliance-compatible compression does not, by itself, preclude manipulation detection, provided the detector operates in a deep differential feature space rather than on hand-crafted texture cues alone.
An extreme-storage variant is now explicit in the 1024-byte study. There, the reference image must remain ICAO-compliant while being compressed to at most 1024 bytes for barcode-based temporary travel documents. In the frontal-only comparison, intended to approximate an ABC-gate scenario with high-quality probes, optimized JPEG AI yields FNMR 0.0718% at FMR 0.0001%, outperforming the uncompressed baseline at 0.4592%. AVIF and WebP also perform well, and the paper states that for comparisons of ICAO-compliant face images only, converting images to grayscale is helpful, whereas for comparisons involving less suitable samples, preserving color is preferable; smoothing and resizing are likewise beneficial (Andreas et al., 29 Jun 2026).
These findings complicate a common assumption that stronger compression necessarily destroys biometric utility. The evidence instead indicates a three-way interaction among codec family, feature representation, and operational comparison regime. This suggests that “ICAO-compliant” should not be conflated with any single storage format or file size, provided the portrait geometry and downstream biometric behavior remain within the required operating envelope.
6. Manipulation threats, integrity protection, and synthetic data
Because ICAO-compliant images are frontal, high-quality, and tightly standardized, they are simultaneously excellent biometric references and attractive attack substrates. Face morphing is the canonical example. The eyebrow-region study reiterates that the face is the primary biometric in passports and that a morphed reference image can allow multiple individuals to authenticate against the same document (Zafar et al., 2023). In response, differential morphing attack detection using deep face representations has shown markedly better robustness than earlier texture or landmark methods: on a database built from ICAO-compliant FERET and FRGCv2 reference images plus unconstrained probes, ArcFace-based differential MAD achieves less than 3% D-EER in many conditions and remains stable under resizing, JPEG 2000, and print-scan post-processing (Scherhag et al., 2020).
More localized or single-image defenses remain relevant but are more fragile. The eyebrow-frequency method, evaluated on print-scan FERET and FRGCv2 morphs, reports D-EER 6.5% on FRGCv2, 22.2% on FERET, and 14.2% on the combined set, indicating that standardized passport-like acquisition can still leave exploitable frequency-domain traces, but that dataset quality and capture pipeline strongly condition performance (Zafar et al., 2023). This suggests that ICAO compliance reduces gross irregularities while not eliminating the need for specialized forensic or differential detectors.
A different line of work treats the ICAO portrait itself as a carrier of integrity signals. Fragile watermarking via deep steganographic embedding stores a hidden marker inside the official photo at issuance time so that post-issuance manipulations alter the recovered marker. The reported experiments show high detection accuracy across multiple manipulation types, including compression, resizing, blur, sharpening, noise, and morphing (Ghiani et al., 18 Apr 2025). The subsequent survey on deep data hiding frames such methods as complementary to PAD because PAD acts only at capture time, whereas embedded signals can support persistent verification without sacrificing ICAO compliance if invisibility, blind extraction, and semi-fragile behavior are maintained (Chivata et al., 26 Aug 2025).
Synthetic data adds a final dimension. ONOT treats ICAO compliance as a hard generation constraint: it begins with 15,000 pseudo-identities and 960,000 generated 512×512 images, but only 4,032 pseudo-classes survive the first ICAO check, about 27%, and only 55, 125, or 255 identities remain after identity-consistency filtering, depending on the threshold (Domenico et al., 2024). This quantifies the difficulty of generating compliant synthetic mugshots rather than merely realistic faces. A related study on StyleGAN and StyleGAN2 finds that 58% of StyleGAN2 images pass ICAO-like pre-selection criteria versus 75% of FRGC images; FaceQnet v1 score distributions are nearly identical between synthetic and FRGC images with KL divergence 0.111, while the ISO-based regressor yields KL divergence 0.444 and SER-FIQ 1.172 (Zhang et al., 2021). Taken together, these results suggest that synthetic ICAO-like imagery is feasible for benchmarking and data augmentation, but that compliance filtering, identity consistency, and demographic balance remain non-trivial bottlenecks.
ICAO-compliant facial images therefore occupy a dual technical position. They are the canonical reference format for interoperable biometric identity, yet their utility depends on a larger ecosystem of quality assessment, storage optimization, attack detection, fairness auditing, and increasingly synthetic-data curation. The accumulated evidence indicates that compliance is best understood not as a static passport-photo template, but as a standards-governed operating point whose security and recognition properties must be continuously engineered.