Human face identification after plastic surgery using SURF, Multi-KNN and BPNN techniques

A facial identification system is capable of recognizing an individual from a video/image series. Human identification systems spot individuals by their facial images. These schemes institute the occurrence of a genuine individual rather than just examining whether a particular detection is valid or not. Identification of an individual is done according to one’s facial morphology. Face acknowledgment is used in various domains such as public retreat systems, marketing, social media, image database investigation, criminal identification, video surveillance and identity verification [23, 26]. Biometric technology has emerged as a promising solution for improving digital security, while fingerprint scans and facial recognition are being used more and more to bring to a standstill theft committed online. Almost all major smart phones are now equipped with some kind of biometric sensor or scanner. Apple just stunned many by introducing face biometrics in place of their fingerprint scanner.

Figure 1 shows a sample image depicting human facial geometry and morphology. Human face detection is not an easy chore because of the deviations in facial gestures, personal looks, poses, clutter, occlusion and lighting effects [18]. Due to the variability in illumination, the quantity of illuminated resources, and the direction of the camera, it is a complicated job to project a forceful face detection scheme in reality with extraordinary identification precision.

A cosmetic therapy is defined as a clinical area of expertise committed towards renovation of facial and bodily imperfections due to birth turmoil’s, shocks, burns, illness, etc. Cosmetic therapy is envisioned to rectify dysfunctional areas of the human body with a reconstructive temperament. Medical/cosmetic alterations are becoming more popular than ever before in the present-day scenario. A considerable rise is witnessed in spending on medical procedures. An impending career lift is one reason as some hope a facial renovation (better personality) could lead to a better job. It has been recently reported that patients feel that their younger appearance will help them to survive in the competitive job market. Since some see it as an investment in their careers, they may use funding options. Funding options have allowed more people to undergo medical therapies who may not have been able to afford it otherwise. As a final point, there is more demand of medical/cosmetic/plastic procedures these days, where a majority of people approve of medical alterations. These surgeries are also popular now among senior citizens. Different medical procedures on the facial region led to transformations in skin consistency and global face appearance [17]. The most obvious benefit of medical surgery is the boost in confidence and self-esteem of an individual. A medical alteration can also help to improve a patient’s physical health and bodily functions.

Facial medical alterations change external face look, which instinctively affects the forcefulness of advent-based face acknowledgment systems. Though these medical surgeries are very useful for beauty and cure purposes, they pose a serious challenge for human identification after surgery as these procedures result into transformation of facial features. There are a variety of local/global medical methods. These include regional operations including Botox injections, laser hair removal, liposuction (treatment for removing extra fat), Blepharoplasty (eyelid surgery), cheek implant surgeries, tummy tucks, otoplasty (reshaping of ears), chin extension and rhinoplasty (restructuring of nose). Face uplift is a global surgery used to smooth out wrinkles and reduce signs of aging. Medical alterations which are local in nature are proficient in giving moderate acknowledgment/recognition consequences because here facial geometry rehabilitated is partial in nature (i.e., type of cosmetic surgery undertaken by individual does not change overall facial appearance such as otoplasty and rhinoplasty), whereas for global alterations, the complete face goes through a rejuvenation (i.e., overall facial appearance is completely changed such as Rhytidectomy and hormone replacement therapies.

Medical surgery supersedes over even the best of recognition algorithms. Solemn crimes are committed by deceivers, by duplicating one’s honest identity thus resulting into great monetary defeats [13]. This is a stern wrongdoing and appeals for widespread responsiveness for investigation in the biometric acknowledgment and authentication dominion for development of a vigorous and unique human identification system [38]. Human recognition is hampered by a variety of factors, including aging effects, position, illumination, expression, disguise, occlusion, presentation attack, type of operation, and dataset effects [42, 49]. The major breakthroughs (motivation and resolved issues) in the manuscript are presented as follows:

The paper is arranged along these lines. After the introduction, the next section comprises of literature review. “Proposed method” describes the proposed procedure with an outline of SURF, Multi-KNN and BPNN. “Results and discussion” is the result section. Deduction and future scope in the facial acknowledgment and authentication domain are discussed in “Conclusion”.

Face recognition after plastic surgery is an interesting research area as the entire facial geometry is transformed depending on the nature of surgery [4, 15, 16, 20, 25, 35‐37]. Local synthetic/plastic therapies reflect minor facial feature alterations while global surgeries change the entire facial geometry and orientation. Face acknowledgment after global synthetic/plastic surgery is difficult as compared to local surgery [8]. Table 1 demonstrates state-of-the-art recognition techniques w.r.t RR, local/global features, texture and 3D aspects. SURF surpassed most of the local feature detection schemes such as Principal Component Analysis (PCA) for dimensionality reduction, Independent Component Analysis (ICA) and Linear Discriminant Analysis (LDA), including PCA hereditary techniques for instance Diagonal/Curvelet/Kernel PCA and DPCA (two-dimensional PCA). It also performed better than other existing recognition methods in literature, namely PCA-DCT, Corr-PIFS, CASNN, and FFNN [7, 10, 21, 29, 31]. Abuzneid et al. [1] proposed a framework for human face recognition using LBPH descriptor, KNN classification and back-propagating neural network. The proposed scheme provided a training dataset with distinct patterns based on the correlation between the original training image samples. The above-mentioned references were in context to non-surgical datasets, whereas the following mentioned references are in context to surgical datasets. Singh et al. [43] offered medical changes as a challenge to facial recognition techniques. To objectively quantify the implementation of face identification techniques on a medically altered dataset that encompasses facial samples for both global and local therapies, the authors presented an experimental study. The outcomes of this study showed that additional study was necessary to create the best facial identification system. Six different facial identification methods were used, including PCA, Fisher Discriminant Analysis (FDA), Local Feature Analysis (LFA), Circular Local Binary Patterns (CLBP), Speeded Up Robust Features (SURF), and Neural Network Design based on Two-Dimensional Log Polar Gabor Transform (GNN). Singh et al. [44] investigated identification rates on the medically altered dataset. These methods were chosen because they offered a combination of feature detection and matching approaches based on emergence, facial characteristics, extractors, and quality. Despite combining the two global and local detection approaches, little equivalence was achieved. Marsico et al. [22] employed association centered facial empathy on brilliance and pose pre-processed illustrations. Bhatt et al. [5] utilized a progressive granular technique with SURF and CLBP feature detectors to process tessellated image samples. Aggarwal et al. [2] used an integrated amalgamation for facial recognition by parts and sparse-based technique. Lakshmiprabha et al. [19] examined Gabor characteristics to use periocular and face biometrics to develop a multimodal biometric system. The proposed methodology had the benefit of not requiring either the training procedure or any contextual information from other datasets. Ibrahim et al. [12] offered comparative examination for numerous recognition schemes applied to pre/post-surgically altered samples. The performance of several Gabor-based facial demonstrations, PCA, KPCA (Kernel PCA), Kernel Fisherface (KFA), and other feature abstraction techniques were compared. With the use of histogram normalization techniques, these strategies were combined. The results showed that gradient-face was the best lighting approach, histruncate was the best histogram normalizing technique, and GKFA was the best feature abstraction method. Mun et al. [24] also examined multimodal biometric systems. This technique matched facial samples before and after medical changes by utilizing several aspects from the face and periocular regions. Local binary patterns were used to extract the features, and PCA was used to reduce the number of dimensions. The categorization process employed Euclidean metrics. The suggested strategy improved the recognition rate by extracting form and textural data. Said et al. [40] estimated geometrical face acknowledgment after medical alterations. Concerned areas of the after-surgery samples were localized and their midpoints were identified. After this, the geometrical distances amid the concerned areas were located to ascertain post-geometrical face vectors. Lastly, a distance classifier (minimal in nature) was employed to equate the before trait vectors of the test illustration image with the after-trait vectors. Pre-feature attributes that gave utmost equivalence and minimum space were measured as the best match. Verghis et al. [48] proposed an algorithm that uses extracted granules from a facial image and is Multi-objective evolutionary granular in nature. First, Gaussian and Laplacian of Gaussian multi resolution image pyramids were created from facial data. Second, facial samples were broken up into horizontal and vertical granules with various sizes. The final step involved extracting data from local face regions. The ability to transition between SIFT and EUCLBP features was made possible through evolutionary feature selection, which also made it possible to encode discriminant information for each face granule. Results from the investigation beat cutting-edge techniques, including a commercial system for equating samples of medically changed faces. In terms of RR, MSE, and F-score, region-based identification techniques produced satisfactory results with levels of moderate to high computation complexity. Bouguila et al. [6] stated that there are limited medical articles discussing the requirement of a robust facial recognition system within the field of plastic surgery. Artificial intelligence is making its way through various aspects in communal and remote areas. The field of plastic surgery takes benefit of the widely available technologies in AI, i.e., deep learning, big data, and NLP; and the recent interests of policy makers in supporting the growth of AI [14]. Algorithms based on AI sub-fields are doing great in the facial recognition domain [3]. Suri et al. [46] proposed an approach which independently learnt shape, color and texture representations from a generic surgical dataset. An unsupervised dictionary learning technique based on a stage-by-stage least angle regression approach was used to learn representations. First, the pre-trained DenseNet for face recognition was used, and then a neural network classifier was developed utilizing features projected into the dictionary space. The only limitation in employing deep models for face recognition after surgery is their inherent complexity and GPU requirements (i.e., large processing power). Scientifically rigorous data are required for biometric identification after plastic surgery. The proposed method achieves a higher Identification Rate and an inferior error rate and computation time (CT) which is an obligation for robust face acknowledgment algorithms.

The proposed technique is explained with a block diagram in Fig. 2. This method demonstrates an enhanced human facial identification system using SURF feature extractor, Multi-KNN, and BPNN network. Our main aim is centered on attaining a robust T-Dataset which helps the BPNN to congregate rapidly with enhanced recognition rate. T-Dataset is fabricated on the correlation among the training samples, and not on the concentration of image samples.

The projected methodology is validated on three datasets, namely AT&T, YALE and the Plastic surgery facial dataset. The projected scheme was applied in three different phases: 70% training/30% testing samples, 90% training/10% testing samples and finally 50% training/50% testing samples.

Using SURF feature extractor, Multi-KNN, and BPNN approaches, we have proposed a system for human face detection. A training database with distinctive arrangements focusing on the correlation among the distinctive training samples has been obtained, thanks to the original SURF extractor and Multi-KNN. Since each distance metric has an advantage over supplemental metrics, this was made possible by combining the distance metrics. We increased recognition rate while lowering computational complexity. Overall, we outperformed current recognition methods. The proposed scheme indicates that using a correlated training database between image samples, higher recognition can be accomplished with conventional feature extraction and dimensionality reduction techniques. The proposed methodology when tested with plastic surgery facial database gives inferior recognition results as compared to non-surgical databases. Local plastic surgeries give satisfactory identification rates, but recognition outcomes for global plastic surgeries fall short of expectations. The significance of machine learning algorithms cannot be foreseen since they use classifiers to teach the unique asset of self-wisdom from prior experiences. Our research’s findings show that the process of facial identification, particularly after surgical modifications, is difficult and necessitates ongoing study. The projected scheme offers moderate computation time in context to other recognition techniques existing in literature. We also plan to test deep neural models for the medically altered facial dataset [32, 33]. Imminent research trends would be to abstract more specific local/global facial traits for a better-quality match between different surgical images of the same person. Employment of a supplementary distance metric can enhance the contest, acknowledgment level and calculation complexity. In addition, face detection done with the proposed method is for still images. This can further be extended for 3D images, i.e., videos and thus the detection will become significantly more complex. The projected methodology surpassed present-day state-of-the-art schemes. Identification after medical changes is a tedious task that necessitates constant investigative efforts. Thus, ongoing research is needed in the field of facial recognition following medical changes. Given that it relates to the fields of biometric authentication, human identification, and how contemporary medical advancements affect face recognition, this research project would greatly benefit society.