Introduction
Against the backdrop of the rapid growth of the global biometrics market, iris recognition technology, with its unique advantages of high precision and high security, is becoming the preferred solution for key scenarios such as financial payment, border security, and smart cities. According to forecasts by market research institutions, the global iris recognition market size will grow from 5.14 billion US dollars in 2025 to 12.92 billion US dollars in 2030, with a Compound Annual Growth Rate (CAGR) of 20.3%.
Figure 1: Global Iris Recognition Market Size Growth Trend (2025-2030 Forecast)
At this critical juncture of industrial transformation, Homsh Technology has successfully achieved a leapfrog upgrade of iris recognition technology from the traditional paradigm to the AI paradigm by virtue of two core invention patents—"An Iris Fast Retrieval System and Method Based on Vector Database" and "An Iris Continuous Feature Encoding Method Based on Deep Neural Networks"—establishing an important innovative position at the forefront of iris recognition technology in China and globally.
Technical Background: Bottlenecks of Traditional Methods and Opportunities in the AI Era
Since the commercialization of iris recognition technology in the 1990s, it has long relied on the IrisCode encoding method based on Gabor filters. This method extracts iris texture features through multi-scale and multi-directional Gabor filters, quantifies them into 2048-bit binary codes, and uses Hamming distance for matching. However, this traditional paradigm faces three core bottlenecks: first, fixed filters cannot adapt to the quality differences of different iris images; second, binarization encoding causes significant information loss, resulting in an Equal Error Rate (EER) of only about 1.75% on the CASIA-Iris-Lamp standard test set; third, the retrieval speed is slow in large-scale databases (over one million level), making it difficult to meet the needs of real-time applications.
Figure 2: Iris Recognition Technology Paradigm Comparison - Traditional IrisCode Encoding vs. Deep Learning Continuous Feature Encoding
With the maturity of deep learning technology and the accumulation of large-scale datasets, iris recognition has ushered in a window of opportunity for paradigm shift from "handcrafted features" to "end-to-end learning". Recent academic research has shown that iris recognition methods based on deep neural networks have demonstrated potential beyond traditional methods. For example, the cutting-edge academic IrisFormer model can achieve an EER of 0.88% on the same dataset. However, how to transform academic achievements into engineering-feasible technical solutions with industrial competitiveness is a common challenge faced by the industry.
Technical Innovation: Two Patents Collaborate to Build a Full-Stack AI Solution
The two core patents released by Homsh Technology this time systematically solve the technical bottlenecks of traditional iris recognition from two dimensions—"feature representation" and "retrieval efficiency", forming a complete technical closed loop from front-end encoding to back-end retrieval.
Figure 3: Homsh Technology's Dual-Patent Collaborative Architecture - End-to-End Iris Recognition System
Patent 1: An Iris Continuous Feature Encoding Method Based on Deep Neural Networks
This patent innovatively combines the EfficientNet-B3 efficient convolutional neural network architecture in the field of computer vision with the ArcFace angular margin loss function in the field of face recognition, realizing end-to-end deep learning encoding of iris features for the first time. Its core innovations include:
1. Adaptive Feature Extraction: Through the compound scaling strategy (three-dimensional balanced expansion of depth, width, and resolution) and efficient MBConv modules (inverted residual structure + squeeze-and-excitation attention mechanism), EfficientNet-B3 achieves high-discriminative feature extraction of iris textures under the constraint of only 12.14 million parameters. Compared with fixed Gabor filters, the model can automatically learn the optimal feature representation.
2. Continuous Feature Encoding: Breaking through the binarization quantization limitation of traditional IrisCode, it outputs 512-dimensional float32 continuous feature vectors with an information capacity of 16,384 bits (8 times that of IrisCode). The feature space is upgraded from a discrete Hamming space to a continuous Euclidean space, enabling more refined similarity measurement.
3. ArcFace Angular Margin Optimization: In the normalized hyperspherical feature space, adding a 10° angular margin forces intra-class aggregation and inter-class separation, reducing the angle between iris feature vectors of the same person and expanding the angle between different people, significantly improving the discriminability of features. Experimental verification shows that compared with the standard Softmax loss, ArcFace reduces EER by 45.4%.
4. Class-Balanced Batch Sampling: To address the problem of uneven sample counts among different individuals in iris datasets, an innovative class-balanced sampling strategy is designed. Each training batch contains 16 classes with 8 samples per class, ensuring that the ArcFace loss function can fully learn inter-class boundaries, accelerating convergence by 30% compared with random sampling.
Patent 2: An Iris Fast Retrieval System and Method Based on Vector Database
This patent applies FAISS (Facebook AI Similarity Search) vector database technology to the field of iris recognition for the first time globally, realizing millisecond-level retrieval in a one-million-person database and providing key technical support for the real-time application of large-scale iris recognition systems. Its core innovations include:
1. FAISS Vector Index Construction: After L2 normalization of the 512-dimensional iris feature vectors extracted by deep learning, the IndexFlatIP index type of FAISS is used for storage. This index type is based on inner product similarity search, which is equivalent to the cosine similarity of normalized vectors. Compared with NumPy brute-force search, it achieves 15.9x CPU acceleration and 75.0x GPU acceleration in a 10,000-person scale database.
2. Intelligent Index Strategy: An innovative multi-level index architecture is designed. Through feature distribution optimization and adaptive clustering, mis-matching is avoided, and flexible recognition modes are supported, significantly improving recognition accuracy and system robustness.
3. Efficient Data Structure Design: The system stores FAISS index files (.index.faiss) and metadata files (.meta.json) separately. The index files are directly mapped to memory for approximate nearest neighbor search, while the metadata files store business information such as personnel IDs, collection times, and device numbers. The query latency is controlled within 8.5 milliseconds (CPU mode).
4. Seamless Integration of Deep Learning Models: The front-end of the system uses EfficientNet-B5 (112MB ONNX) for iris segmentation to extract the region of interest; the back-end uses EfficientNet-B3+ArcFace (44MB ONNX) for feature extraction. The entire process is end-to-end optimized from image input to retrieval result output, supporting both CPU and GPU inference modes and adapting to various deployment scenarios such as edge devices and servers.
Technical Indicators: Reaching World-Class Levels
Rigorous tests on the international standard iris dataset CASIA-Iris-Lamp (573 people, 11,845 images) show that Homsh Technology's dual-patent solution has achieved the following breakthrough indicators:
Figure 4: Iris Recognition Performance Comparison (CASIA-Iris-Lamp Dataset)
1. Equal Error Rate (EER): 0.70%. Compared with the traditional Gabor+Hamming distance method (1.75% EER), the error rate is reduced by 60%; compared with Homsh Technology's previous EfficientNet-B3 baseline solution (2.66% EER), the error rate is reduced by 73.7%; compared with the cutting-edge academic IrisFormer model (0.88% EER), the performance is improved by 20.5%, establishing a world-class leading position in the industry.
2. Recognition Accuracy (AUC): 99.97%, indicating that a very high correct recognition rate can be maintained even at an extremely low false recognition rate.
3. Retrieval Speed: In a 10,000-person scale database, the average retrieval latency is 8.5 milliseconds in FAISS CPU mode with a throughput of 117.6 QPS; the retrieval latency is 1.8 milliseconds in GPU mode with a throughput of 555.6 QPS. Compared with traditional NumPy brute-force search, it achieves 15.9x and 75.0x acceleration respectively, fully meeting the needs of real-time applications.
4. Model Efficiency: The EfficientNet-B3 feature extraction model has only 12.14 million parameters, with an ONNX inference time of 8 milliseconds (CPU) and a memory footprint of 1.8GB, supporting deployment on edge devices and mobile terminals; through INT8 quantization, the model size can be further compressed to 11.2MB, the inference time reduced to 5 milliseconds, and the memory footprint reduced to 0.5GB.
Industry Leadership: Dual Innovations from ASIC Chips to AI Paradigms
Homsh Technology has unique technical accumulation and innovative genes in the forefront of China's iris recognition technology. As early as before 2020, the company successfully developed the world's first ASIC chip dedicated to iris recognition, breaking through the hardware acceleration bottleneck of iris recognition algorithms, increasing the recognition speed to the millisecond level, and laying a hardware foundation for the large-scale commercialization of iris recognition technology. This innovation has given Homsh Technology a first-mover advantage in the industrialization process.
Entering the AI era, Homsh Technology keenly captured the opportunity of deep learning technology to restructure the iris recognition paradigm, resolutely invested in R&D resources, and achieved a paradigm upgrade from "traditional signal processing" to "end-to-end deep learning" in two core dimensions: encoding methods and retrieval systems. The dual-patent solution released this time not only achieves a world-class EER level of 0.7% in technical indicators but also, more importantly, realizes the global pioneering application of FAISS vector database in the field of iris recognition, filling the gap in this technical route. This marks that Homsh Technology has completed the strategic transformation from a "chip innovator" to an "AI paradigm leader", establishing a technological commanding height in the era of intelligent iris recognition.
Potential Applications: Empowering Intelligent Upgrade in Multiple Fields
With its technical advantages of high precision, high speed, and easy deployment, Homsh Technology's dual-patent solution can be widely applied in the following scenarios:
Figure 5: Application Scenarios of Homsh Technology's Dual-Patent Solution
Financial Payment
Deploying iris recognition on bank ATMs and mobile payment terminals, the ultra-low EER of 0.7% ensures fund security, the 8-millisecond recognition speed provides a smooth user experience, and the single-eye mode supports users wearing glasses.
Border Security
Deploying large-scale iris recognition systems at airports and ports, the FAISS vector database supports millisecond-level retrieval in a one-million-person database, and the multimodal fusion strategy further improves accuracy, effectively preventing identity fraud.
Smart Parks
Deploying iris access control in enterprise parks and government agencies, the INT8 quantized model supports local deployment on edge devices (access control machines, turnstiles), enabling real-time recognition without networking and ensuring data privacy.
Healthcare
Integrating iris recognition into hospital HIS systems to accurately associate patient identities with electronic medical records, avoiding confusion caused by the same name and improving medical safety; establishing a unique biometric ID in newborn management to prevent baby abduction.
Public Security
Deploying iris recognition in urban monitoring systems, combined with long-distance iris collection equipment, to realize early warning of key personnel monitoring and control. The GPU inference mode supports high-concurrency real-time analysis.
CEO's Remarks: Dr. Yi Kaijun, CEO
Dr. Yi Kaijun, CEO of Homsh Technology, stated in an interview: "The successful R&D of these two patents is the crystallization of Homsh Technology's more than a decade of technical accumulation and continuous investment in innovation. We deeply understand that in the highly competitive field of biometrics, only by mastering core technologies can we remain invincible. From the ASIC chip innovation before 2020 to today's dual breakthroughs in deep learning + vector database, Homsh Technology has always adhered to the in-depth integration of cutting-edge technology and industrial needs. The 0.7% EER indicator is not just a number; it represents the optimal balance between 'security' and 'usability' achieved by the system. For key scenarios such as finance and security checks, this means higher security guarantees and a better user experience."
"More importantly, we are the first in the world to introduce FAISS vector database technology into the field of iris recognition. This innovation opens up new possibilities for the real-time application of large-scale iris recognition systems. In the future, we will continue to deepen our efforts in the field of AI + biometrics, promote the application of iris recognition technology in more scenarios, and contribute Homsh's strength to the construction of a smart society. Innovation is endless, and Homsh Technology will continue to lead the technological progress of the industry."
Outlook: The Future of Intelligent Iris Recognition
With the continuous evolution of AI technology and the improvement of infrastructure such as 5G and edge computing, iris recognition is moving from "specialized scenarios" to "inclusive applications". Homsh Technology's dual-patent solution, with its outstanding technical performance and engineering capabilities, is fully prepared to meet the market explosion in the next decade. The company will continue to invest in R&D resources and make continuous innovations in directions such as multimodal fusion (iris + face + fingerprint), liveness detection, and privacy computing, contributing core technical strength to building a safer, smarter, and more convenient digital society.
About Homsh Technology
Homsh Technology is a leading iris recognition technology provider in China, focusing on the R&D and industrialization of iris recognition algorithms, chips, and systems. The company holds a number of core technology patents including the world's first ASIC chip dedicated to iris recognition, and its products are widely used in fields such as finance, security, and healthcare.
