A robust, low-cost approach to Face Detection and Face Recognition

In the domain of Biometrics, recognition systems based on iris, fingerprint or palm print scans etc. are often considered more dependable due to extremely low variance in the properties of these entities with respect to time. However, over the last decade data processing capability of computers has increased manifold, which has made real-time video content analysis possible. This shows that the need of the hour is a robust and highly automated Face Detection and Recognition algorithm with credible accuracy rate. The proposed Face Detection and Recognition system using Discrete Wavelet Transform (DWT) accepts face frames as input from a database containing images from low cost devices such as VGA cameras, webcams or even CCTV’s, where image quality is inferior. Face region is then detected using properties of Lab* color space and only Frontal Face is extracted such that all additional background is eliminated. Further, this extracted image is converted to grayscale and its dimensions are resized to 128 x 128 pixels. DWT is then applied to entire image to obtain the coefficients. Recognition is carried out by comparison of the DWT coefficients belonging to the test image with those of the registered reference image. On comparison, Euclidean distance classifier is deployed to validate the test image from the database. Accuracy for various levels of DWT Decomposition is obtained and hence, compared.

💡 Research Summary

The paper presents a low‑cost, robust pipeline for face detection and recognition that works with low‑resolution video sources such as VGA cameras, webcams, and CCTV systems. Recognizing that many biometric modalities (iris, fingerprint, palm‑print) require high‑quality data and user cooperation, the authors argue that a real‑time, fully automated facial solution is needed for practical security applications.

The system consists of four main stages. First, the input image is converted from the standard RGB space to the perceptually uniform CIE Lab* space. By thresholding the a* and b* channels the algorithm isolates skin‑like regions, then uses the presence of “holes” (dark patches corresponding to eyes, nostrils, mouth) to confirm that the region is a frontal face. This approach is chosen over Gabor filters, neural‑network skin classifiers, or template matching because it is computationally cheap and does not require training data.

Second, the detected face is cropped, resized to 128 × 128 pixels, converted to grayscale, and intensity‑normalized so that the average brightness matches that of the reference images. This step mitigates illumination variations between capture sessions.

Third, a discrete wavelet transform (DWT) is applied to the whole normalized face. The authors experiment with multiple decomposition levels (1–4) using a basic wavelet (e.g., Haar). Only the approximation (low‑frequency) coefficients are retained as the feature vector; detail coefficients are discarded to reduce noise sensitivity and computational load. The resulting vectors are low‑dimensional yet capture the overall facial shape.

Finally, recognition is performed with a simple Euclidean‑distance nearest‑neighbor classifier. The test image’s feature vector is compared against all stored reference vectors, and the identity associated with the smallest distance is returned.

The experimental protocol uses a small proprietary dataset: 50 subjects, each providing five training (reference) images and five test images captured under varying lighting conditions with low‑resolution devices (as low as 320 × 200). Accuracy is reported for each DWT level, showing a modest increase with deeper decomposition but also a rise in processing time. The authors claim that the method achieves “credible” accuracy while requiring far less computation than Gabor‑filter, deep‑learning, or template‑matching approaches, making it suitable for embedded or resource‑constrained platforms.

However, several limitations are evident. The skin‑color detection in Lab* space may be sensitive to ethnicity and extreme lighting; the paper does not provide quantitative analysis of false‑positive/false‑negative rates across diverse skin tones. By discarding high‑frequency detail, the system may struggle with fine‑grained discriminative cues needed for high‑security verification. The Euclidean‑distance classifier, while fast, lacks robustness to intra‑class variability and may be outperformed by modern metric‑learning embeddings (e.g., FaceNet, ArcFace). Moreover, the dataset is small and lacks the variability present in public benchmarks such as LFW, MegaFace, or IJB‑C, so the reported performance cannot be directly compared to state‑of‑the‑art methods. No runtime measurements or memory footprints are provided, leaving the claim of real‑time suitability unverified.

In summary, the paper contributes a straightforward, low‑complexity face detection‑recognition pipeline tailored for inexpensive, low‑resolution cameras. It demonstrates that DWT‑based low‑frequency features combined with Lab* skin segmentation can yield reasonable recognition rates without heavy computational resources. Future work should address cross‑ethnicity skin detection robustness, incorporate high‑frequency or learned features to improve discriminability, evaluate on large, diverse benchmarks, and provide detailed performance metrics on target hardware.