Attribute-Aware Face Aging With Wavelet-Based Generative Adversarial Networks

“To be specific, we employ an hourglass-shaped fully convolutional network as the generator, which has achieved success in previous image translation studies [6, 29].”

“All face images are cropped and aligned according to the five facial landmarks detected by MTCNN [27].”

“To be specific, we utilize the VGG-Face descriptor [14], denoted by φ, to extract the identity related semantic representation of a face image.”

“Unlike regular GANs [5], we adopt least square loss instead of negative log likelihood loss for that margins between generated samples and the decision boundary in the feature space are also minimized, which further improves the quality of synthesized images [12].”

“Unlike regular GANs [5], we adopt least square loss instead of negative log likelihood loss for that margins between generated samples and the decision boundary in the feature space are also minimized, which further improves the quality of synthesized images [12].”

“[28] proposed a conditional adversarial autoencoder (CAAE) to achieve age progression and regression by traversing in low-dimensional manifold.”

“been proposed to address this problem in the last two decades [8, 20, 19, 21, 7].”

“Data samples in MORPH are color images of near-frontal faces exhibiting neutral expressions under uniform and moderate illumination with simple background.”

“To suppress such undesired changes of semantic information during the aging process, many recent face aging studies attempt to supervise the output by enforcing identity consistency [28, 1, 25, 9].”

“For CACD, since face images with race other than ‘white’ only takes a small portion of the entire dataset, we only select ‘gender’ as the attribute to preserve.”

“been proposed to address this problem in the last two decades [8, 20, 19, 21, 7].”

“With the success of GANs [5] in generating visually appealing images, many efforts have been made to tackle the problem of face aging using GAN-based framework [28, 25, 9, 17, 24, 10].”

“[22] modeled the translation of facial appearance by revised cardioidal strain transformation.”

“been proposed to address this problem in the last two decades [8, 20, 19, 21, 7].”

“In recent years, deep generative models with temporal architectures are adopted to synthesize images of elderly faces [23, 3, 4].”

“However, our method further generates face images of higher resolution (2×) with enhanced details compared to GLCA-GAN, and reduces ghosting artifacts in the results compared to PAG-GAN (e.g. finer details of hair and beard).”

“We choose ‘gender’ and ‘race’ to be the attributes that are required to be preserved, since these two attributes are guaranteed to remain unchanged during natural aging process, and are relatively objective compared to attributes such as ‘attractive’ or ‘chubby’ used in popular facial attribute dataset CelebA [11].”

“Face aging, also known as age progression [16], aims at rendering a given face image with aging effects while still preserving personalized features.”

“These four methods are (from left to right): CONGRE [18], HFA [26], GLCA-GAN [9], and PAG-GAN [25].”

“With the success of GANs [5] in generating visually appealing images, many efforts have been made to tackle the problem of face aging using GAN-based framework [28, 25, 9, 17, 24, 10].”

“Following the convention in [25, 9], we divide the face images into four age groups, i.”

“In recent years, deep generative models with temporal architectures are adopted to synthesize images of elderly faces [23, 3, 4].”

“These four methods are (from left to right): CONGRE [18], HFA [26], GLCA-GAN [9], and PAG-GAN [25].”

“With the success of GANs [5] in generating visually appealing images, many efforts have been made to tackle the problem of face aging using GAN-based framework [28, 25, 9, 17, 24, 10].”

“been proposed to address this problem in the last two decades [8, 20, 19, 21, 7].”

“Note that this is different from [9], since wavelet coefficients are only used for discrimination in our work and no prediction or reconstruction is involved.”

“In recent years, deep generative models with temporal architectures are adopted to synthesize images of elderly faces [23, 3, 4].”

“Face generation [332]–[340] is somewhat easy because there is only one class of object.”

“Generative Adversarial Networks (GANs) have introduced numerous applications across various domains, including image blending [67], 3D object generation [68], face aging [69], medicine [70, 71], steganography [72], image manipulation [73], text transfer [74], language and speech synthesis [75], traffic control [76], a…”

“Face aging: Age-CGAN [49], CAAE [50], IP-CGAN [51], Wavelet-GANs [52, 53]”

“Liu et al. [89] introduced an attribute-aware face aging model with wavelet-based GANs for aged face image synthesis.”

“d synthesis. All of the above-mentioned methods operate at resolution of at most 256x256 when applied to human faces. Gender swap is one of well-known tasks of unsupervised image-to-image translation [10,11,37]. Face aging/rejuvenation is a special task which gets a lot of attention [59,49,18]. Formulation of the prob…”

“e restoring image details. Wavelet-based Methods. Wavelet-based methods have been explored in some computer vision tasks, including classication [7,38,29,57], network compression [28,11], face aging [36], super-resolution [19,35], style transfer [63], etc. Fujieda et al. [7] proposed wavelet CNNs that utilize spectral…”

“To investigate the effects of wavelet filters over the inpainting performance, we evaluated another two types of wavelet filters including db2 from the Daubechies wavelet family and bior2.”

“ynthesize the aged face. Wang et al. [44] added facial feature loss as well as age classication loss to enforce the output image to have the same identity while still progressing the age. Liu et al. [29] added gender and race attributes to their GAN architecture to help avoid biases in the training data, they also pro…”

“Recently, generative adversarial networks (GANs) [12] have been successfully used to enhance the image quality of synthesized faces [24, 27, 51, 56, 58]; they typically use the one-hot encoding to specify the target age group.”

“Following [15, 17, 19], we divide all ages into N nonoverlapping age groups, where N = 4 in this paper.”

“In contrast to traditional T2I tasks which generate flowers and birds, Text-to-Face (T2F) is more challenging and rarely investigated because face images have many fine-grained features [6, 20, 41, 42, 47].”

“Unlike the latest works on face aging [4], [5], [6], [7], [8], our unconditioned discriminator is used to distinguish between real and manipulated images.”

“face hallucination [41] and face aging [42].”

“Wavelet decomposition based approaches played crucial role in various computer vision tasks including GAN-inversion [33, 45], generative modeling [14, 49, 56], face-aging [5, 34], video compression [59], medical and thermal imaging [61, 66].”

“In this case, visual attribute vectors are regarded as the conditional information and embedded into both the generator and discriminator, encouraging synthesized images to be faithful to the visual attributes of the corresponding inputs [59], [60].”

“been successfully used to enhance the image quality of synthesized faces [10], [11], [12], [13], [14], which typically use one-hot encoding to specify the target age group.”

“Others [21, 32] have augmented discriminators by providing them with not only an image, but also frequency sub-bands derived from a wavelet packet transform.”

“• The S-module is orthogonal to the existing methods, therefore can be used as a plug-in module in many recent methods with a transformation network such as [17, 42, 19], to reduce their computational consumption as well as enable the continuous aging, while still keeping their own advantages.”

“The application fields based on WT and deep learning methods are image classification [10, 23], computer vision [4, 17], texture classification [6], etc.”

“Similarly, wavelet information is used in [28, 27], but in a multi-scale manner in the discriminator.”

“For example, they have been used for wavelet-based feature pooling operations [14,30,57], for the improvements on style transfer [13, 61], for denoising [29], for medical analysis [24], and for image generation [21,32,42,56,56].”

“Solutions to such challenges including partial or occluded face detection [1], 3D facial landmark detection [2], face frontalization [3], facial de-makeup [4], heterogeneous face recognition [5], pose-invariant face recognition [6], kinship identification [7], liveness detection [8], facial aging [9] and age estimatio…”

“Compared to our previous work in [20], this paper has the following extensions: 1) attention mechanism is introduced to help improve the visual quality of generation results via restricting modifications to image regions closely related to age progression; 2) generalization ability of the proposed method is investigat…”

“Given that people from different demographic groups have different aging processes, Liu et al . designed an attribute-aware face aging model [26].”

“Generative Adversarial Network (GAN) [9] has been successfully applied to face aging and made impressive re-sults [8, 21, 22, 24, 26, 25, 23, 28, 1, 10].”

“We observe that the generated ages by DAAE are closer to the real data than by CAAE [40] as well as GLCA-GAN [18], and comparable to [35, 20].”

“Authors of [21] improved image quality of synthesized images by using a wavelet packet transformation and multiple facial attribute encoding.”

“W Applications include Convolutional Neural Network (CNN) augmentation [58], style transfer [63], image denoising [30, 44], image coloring [29], face aging [31], video enhancement [54], face super-resolution [18], and generative machine learning [11, 12, 65, 39].”

“With advances in Generative Adversarial Networks [Goodfellow et al., 2014] in recent years, tremendous progress has been made in face portrait editing [Yang et al., 2018; Choi et al., 2018; Liu et al., 2019].”

“With the development of deep learning [9, 23, 26, 27, 45], an increasing number of works [2, 3, 28, 44, 47, 49, 55] put their efforts on exploring better features for simple matching algorithms (e.g. Hungarian matching).”

“[45], who also operate within the frequency domain to extract textual features at multiple scales more effective.”

“Learning in the wavelet domain has the advantage of explicitly dealing with signals in different frequency sub-bands, and it has been applied to some high-level vision and low-level vision problems, such as classification [16], [20]–[22], face aging [23], style transfer [24], image denoising [17], [25], [26], image de…”

“Unlike [28] which is hard-coded, limited to a single attribute, our method is automatic, and supports arbitrary multiple attributes.”

“During training, we adopt an architecture similar with [6] which is shown in Fig.”

“Most cGANs feed the conditional variable y into the discriminator by concatenating y to the input or to some feature vectors [14,18,25,21].”

“In fact, identity features have been widely used to represent identity information in many tasks, such as face clustering [67] and face aging [42].”

“we explain the overall loss used to train our method. A. Problem Formulation Since the transformation is conducted among age categories rather than specific ages, following the prior work in [3], [5], [52], we divide the data into four categories according to the following age ranges: 30 , 31 40, 41 50, and 51+. Each c…”

“Liu et al. [24] propose a wavelet-based discrimi-nator for generating fine-grained details of aging effects.”

“PAG-GAN and WB-GAN produce excellent results in all cases.”

“Some methods [10] could not synthesize face images of multiple age groups.”

“25, 27, 29, 41–43] to estimate identity preservation objectively, the online face analysis tool Face ?? is used.”

“Under the context of 2D-aware aging, Generative Adversarial Networks (GANs) [13] are widely employed for face aging and have demonstrated impressive results, as shown in [1, 16, 18, 26, 27, 30, 34, 39, 41, 43, 44].”

“The visualization can be potentially improved by many recent generative works [57].”

“But current face aging method used ‘‘one-shot” and achieve the state-of-the-art (Despois et al., 2020; Liu et al., 2019; Fang et al., 2020).”

“To evaluate the face aging accuracy of the three methods, following the works by [20], [21], we estimated the ages of the generated face images via the Identity Preservation.”

“Over the past decades, wavelet-based methods have been explored in many computer vision tasks, including image classification [36], face aging [37], style transfer [38], etc.”

“For instance, Liu et al. [36] employed Face ++ APIs to estimate age and analyzed the distribution of estimated ages to evaluate age accuracy.”

“Thus, GANs have been applied in image combination [5], image manipulation [6], image enhancement and inpainting [7], Content Based Image Retrieval (CBIR) [8], changing faces over time [9], face completion [10], human poses [11], face expression recognition based on generative adversarial networks [12], object detectio…”

“[32] incorporates a wavelet packet transform module to improve the visual fidelity of generated images by capturing age-related texture details at multiple scales in the frequency space.”

“Liu et al. [14] utilized wavelet-based GANs to perform attribute-aware facial aging.”

“The one-step approach to facial aging is still a top priority, and there are still many challenges to producing synthetic faces for certain age groups using only one training course[33], and the current face aging method used “one-shot” and achieves state-of-the-art[34][35][36].”

“To address the matching ambiguity between young and aged faces inherent to the unpaired training data, Liu et al. [4] propose an attribute-aware facial aging model with wavelet-based GAN (WL-GAN).”

“In current deep learning, wavelet transform is usually used as image pre‐processing or post‐ processing.(38,39) The discrete wavelet transform (DWT) can decompose an image into four components by sampling horizontal and vertical channels using sub‐band filters.”

“In recent years, great progress has been made with GANs in the field of face generation [32], [37], [13], [6], [17], [16], [4], [18].”

“Liu et al. [36] proposed a framework based on Generative Adversarial Networks (GAN) to embed facial attribute vectors in the generator and discriminator.”

“Liu et al. [49] focused on age attributes, presenting a wavelet-based GAN model that captures age-related texture details across multiple scales in frequency space, effectively addressing challenges in face aging with unpaired datasets and mitigating unnatural facial attribute changes.”

“In future work, facial morphing techniques [65] are desirable to integrate facial characteristics.”

“…[Kang et al. , 2018; Liu et al. , 2020b], super resolution [Liu et al. , 2018], classification [Li et al. , 2020; Liu et al. , 2020a], segmentation [Li and Shen, 2020], facial aging [Liu et al. , 2019], style transfer [Yoo et al. , 2019], remote sensing image processing [Duan et al. , 2017], etc.”

“Our objective is capturing high-level information and local details for Chinese paintings, First, we utilize haar wavelet transform algorithm [17], [16], [7] to decompose specific images into various frequencies due to its in dealing with approximate or detail information of images.”

“In addition, facial attributes information [54] and spatial attention mechanisms [55] are incorporated”

“Existing GAN-based works [21, 22, 23, 24, 25, 26] focus on simulating the human face aging process.”

“To alleviate this problem, existing face aging studies [4, 5, 6, 7, 10, 11] group several continuous ages into discrete age groups (e.”

“Most prior works [8, 122, 103, 70, 47, 77, 71, 5, 15], especially those based on Generative Adversarial Networks (GANs) [23], fall under this category.”

“Other approaches [11], [24], [25] embed age attributes into latent variables using autoencoder-style generator structures and enforce decoupling constraints between age attributes and other attributes, achieving high-quality age transformation effects through adversarial learning.”

“The inverse discrete wavelet transform (IDWT) can reconstruct the original data from the DWT output (Huang et al. 2017; Liu, Li, and Sun 2019).”

“Our method only needs one model to preserve the facial attributes for na age groups, while the PAG-GAN and Attribute-aware GAN need to train na(na+1) models.”

“It is noteworthy that despite [20], [21], [22], [40] also incorporated gender or race attributes; however, they explicitly mentioned in their papers that their objective was to enhance attribute preservation, not to manipulate aging.”

“For example, face aging[45, 37, 18, 19] is a kind of face attribute editing application that naturally alters the apparent age of a face.”

“Liu et al. [110] focused on age attributes, presenting a wavelet-based GAN model that captures age-related texture details across multiple scales in frequency space, effectively addressing challenges in face aging with unpaired datasets and mitigating unnatural facial attribute changes.”

“Indeed, a significant number of studies in the convolutional neural network [11, 29, 34, 47, 56] and the generative domain alike [17, 30, 53, 65]”

“Following [6, 7, 8, 9], we divide all ages into N = 4 nonoverlapping age groups; i.”

“A wide variety of studies [3], [14], [15], [16], [17], [18], [19], [20], [21], [22] have extensively explored the use of GAN models to achieve greater realism in age transformation.”

“At-tributes like age, identity, and expression are often entangled (Liu, Li, and Sun 2019) , so modifying one may unintentionally affect others.”

“Most of these works have focused on the direct application of first-generation WT to CNNs.”

“Previous work targets on changing some predefined attributes, such as the hair color [6] and the age [19], or editing attributes by shape changes [14].”