Search

Mental disorders affect nearly one billion people globally, 94% of whom are undiagnosed and untreated due to an acute shortage of trained clinicians. In response to this crisis, this study introduces mental disorder scan (MDscan), a novel artifact for screening ten mental disorders using data from the SCL-90-R mental disorder screening instrument, an explainable artificial intelligence approach, and our own ShapRadiation algorithm. MDscan converts 90 mental health indicators for each patient into an easily interpretable diagnostic image for mental disorders, similar to radiological images, and explains which indicators contributed to that prediction, increasing clinicians' ability to screen more patients in less time. A field evaluation with clinical data shows that MDscan has high classification accuracy, with average F1 scores between 0.77 and 0.94, compared against prerecorded ground truth. Furthermore, unlike traditional black-box models, MDscan's transparency and explainability can help enhance trust in artificial intelligence (AI) applications for clinical use. [ABSTRACT FROM AUTHOR]

Diffusion models have achieved remarkable results in image generation. However, due to the slow convergence speed, room for enhancement remains in existing loss weight strategies. In one aspect, the predefined loss weight strategy based on signal-to-noise ratio (SNR) transforms the diffusion process into a multi-objective optimization problem. However, it takes a long time to reach the Pareto optimal. In contrast, the unconstrained optimization weight strategy can achieve lower objective values, but the loss weights of each task change unstably, resulting in low training efficiency. In addition, the imbalance of lossy compression and semantic information in latent space diffusion also leads to missing image details. To solve these problems, a new loss weight strategy combining the advantages of predefined and learnable loss weights is proposed, effectively balancing the gradient conflict of multi-objective optimization. A high-dimensional multi-space diffusion method called Multi-Space Diffusion is also introduced, and a loss function that considers both structural information and robustness is designed to achieve a good balance between lossy compression and fidelity. The experimental results indicate that the proposed model and strategy significantly enhance the convergence speed, being 3.7 times faster than the Const strategy, and achieve an advanced FID = 3.35 score on the ImageNet512. [ABSTRACT FROM AUTHOR]

The technology of garment relighting carries substantial research significance in the domains of online garment sales virtual fitting and personalized customization. Moreover relighting methods based on implicit 3D models have garnered considerable attention in the fields of computer vision and computer graphics. However existing scene relighting techniques face inherent challenges when applied to garment datasets such as texture noise and relighting distortion. To address these shortcomings this paper proposes an innovative garment relighting method that operates in the feature tensor to effectively mitigate these issues. The relighting process involves incorporating the relighting component into the tensorial radiance fields to jointly optimize the feature space. To simulate the direct ambient light and the indirect light reflected between garment surfaces Spherical Gauss and MLP techniques are employed to construct an incident light field. Additionally a gradient-guided smoothing strategy is utilized to optimize the extraction of parameters from the bidirectional reflectance distribution function model which are derived from the feature tensor. Finally the garment relighting image is rendered by combining the incident light field the bidirectional reflection distribution model and the feature tensor using the simplified reflectivity equation. This article presents experimental results on three garment datasets comparing them with advanced methods such as Physg and InvRender. The results demonstrate that our method achieves an average improvement of about 4. 549% in generating garment images from novel view and approximately 9. 922% in generating garment images under relighting conditions as evaluated using three indicators. The article visually demonstrates the effectiveness of our proposed method in reducing texture noise and reillumination distortion. Ablation experiments are also conducted examining the impact of gradient-guided smoothing strategies and the use of single or multiple addition feature tensor for garment feature storage. The article shows that using multiple addition feature tensor does not enhance the quality of garment image generation but increases the training time. By comparison using addition single feature tensor achieves a minimal improvement of only 0. 327% but significantly increases the training time. Our method which employs a single feature tensor significantly shortens training time and improves the quality of garment relighting images by approximately 8. 870% compared to the other models. The article compares the experimental results of different lighting strategies DL DL + SGID DL + MLPID DL + SGID + Vis and DL + MLPID + Vis in garment image rendering. The results indicate that the combination of indirect light direct light and visibility achieves the best generation outcomes. The article compares experimental results obtained by employing different gradient descent strategies GD GD + RSL and GD + GGS in garment image rendering. The results indicate that our proposed gradient-guided smoothing strategy enhances the quality novel view garment images by approximately 4. 583% and relighting garment images by about 6. 096% compared to traditional relative smoothing loss methods. This paper introduces a garment relighting model based on Tensorf for 3D garment relighting. The relighting module is integrated into Tensorf which encompasses the incident light field combined with indirect light the BRDF model based on the Disney principle and the simplified reflectance equation. As a result garment relighting from any perspective is achieved. The experimental results demonstrate that the proposed garment relighting model effectively leverages the feature tensor from Tensorf. The introduction of gradient smoothing loss contributes to the improved accuracy of BRDF parameters reduces texture noise commonly encountered in existing methods applied to garment datasets and enhances the accuracy of relighting. The evaluation indexes indicate that the model produces superior results compared to existing advanced methods for generating novel view images and relighting outcomes. However it is worth noting that the surface reconstruction in this method relies on the tensor radiation field which may result in rendering points with low effectiveness leading to errors in the reconstruction of hollow areas on the surface. Future work will address this issue accordingly. [ABSTRACT FROM AUTHOR]

Exemplar-free class incremental learning is a highly challenging setting where replay memory is unavailable. Methods relying on frozen feature extractors have drawn attention recently in this setting due to their impressive performances and lower computational costs. However, those methods are highly dependent on the data used to train the feature extractor and may struggle when an insufficient amount of classes are available during the first incremental step. To overcome this limitation, we propose to use a pre-trained text-to-image diffusion model in order to generate synthetic images of future classes and use them to train the feature extractor. Experiments on the standard benchmarks CIFAR100 and ImageNet-Subset demonstrate that our proposed method can be used to improve state-of-the-art methods for exemplar-free class incremental learning, especially in the most difficult settings where the first incremental step only contains few classes. Moreover, we show that using synthetic samples of future classes achieves higher performance than using real data from different classes, paving the way for better and less costly pre-training methods for incremental learning. [ABSTRACT FROM AUTHOR]

The human face, being one of the most prominent physical features, plays a crucial role in appearance description and recognition. Consequently, text-to-face synthesis has garnered increasing interest in the research community, with applications in criminal investigation, image editing, and more. Compared to text-to-image synthesis, generating facial images from text requires more specialized knowledge due to the subjectivity and diversity of facial descriptions, which involve more fine-grained appearance features. In this paper, we propose a text-to-face synthesis model based on Facial Landmarks Prediction (FLP-GAN). Specifically, we design two foundational submodules to facilitate the generation task. First, a co-attention mechanism is employed to pretrain the image and text encoders to extract features related to facial information. Second, a facial landmarks prediction model is proposed to generate face segment maps based on descriptive text, providing facial semantic prior knowledge for the subsequent face synthesis process. Conditioned on the semantic features obtained from the submodules, we construct the text-to-face synthesis model, which incorporates a memory network and a segment fuse layer to highlight important text information and refine the features. Additionally, a multi-stage refinement process is designed to generate high-resolution face images. Experimental results on the Face2Text dataset demonstrate that our FLP-GAN model outperforms the state-of-the-art methods in both qualitative and quantitative evaluations. Specifically, our model achieved a 22.7% improvement in Fréchet FaceNet Distence compared to the SOTA models. [ABSTRACT FROM AUTHOR]

In the medical sector, diagnostic technology-related progress is often hindered by data isolation and stringent privacy laws, posing obstacles for institutions that lack extensive disease data. This scarcity impedes the development of precise diagnostic models and reliable auxiliary tools. To address these challenges, we introduce the horizontal federated data augmentation model for medical assistance (HFDAM-MA), a novel approach designed to address the complexities of data scarcity. Our model addresses the limitations of traditional generative adversarial networks (GANs), which often rely on the independent and identically distributed (IID) assumption during training (a condition that is rarely satisfied in real-world medical data scenarios) and face computational challenges in healthcare settings. The HFDAM-MA leverages federated learning (FL) principles to enable non-IID collaborative training across multiple medical institutions. This approach alleviates the data collection pressure at individual sites and ensures the privacy of sensitive medical information. A central node orchestrates the distribution of a unified GAN model to local sites, where it operates in conjunction with two convolutional neural networks (CNNs) to generate synthetic medical images and corresponding labels. Extensive experimental results underscore the effectiveness of our model. As participation increases, we observe a substantial improvement in the diagnostic accuracy of the global model. Moreover, the performance of the local models is bolstered, and the diversity of the generated data is expanded, offering a robust solution to the challenges of data privacy, imbalanced data, and insufficient labeling that are prevalent in the medical sector. [ABSTRACT FROM AUTHOR]

Due to its good performance, synthetic aperture radar (SAR) is gradually regarded as a favorable tool to solve the target detection task, especially in marine target detection. Maritime target detection models always need a lot of SAR images for learning, but in fact the use of SAR images has limitations so the acquisition of SAR images is a problem. In addition to this, the labeling of the target is also a time consuming and labor intensive process. In order to solve these problems, we propose a location-based SAR generation model, referred to as the LocSAR, which takes the target position and category information as inputs and outputs location-specific high-quality complex SAR images. The model consists of four parts: a cascade Encoder-Decoder, a scene fusion module, a feature fusion module and a discriminator. We use the feature fusion module to fuse the features of neighboring resolution images and introduce the scene fusion module for semantic information learning. The Encoder-Decoder implements feature extraction and reconstruction, while the discriminator is used for adversarial training. We also use gradient variance loss to improve the quality of the generated images. We updated the existing data set HRSID with the addition of land and sea labels and conducted experiments on this. Extensive experiments show that our model is excellent to classic image generation models. We have also designed a simple interface to assist in image generation and labeling, which can be directly applied to the other model. [ABSTRACT FROM AUTHOR]

Medical image enhancement is considered a challenging image-processing framework because the low quality of images resulting after acquisition and transmission seriously affects the clinical diagnosis and observation. In order to improve the medical image visual quality, a novel medical image enhancement algorithm that is based on contrast adaptive histogram equalization and pelican optimization algorithm is proposed in this work. The estimation process using our proposed model improves the efficiency of the operation and provides superior results in terms of image quality and contrast. There are three steps in the enhancement process. The primary step includes medical image generation using a Text-to-image generative model. Secondly, the estimation of the clip-limit, which controls the enhancing performance. Finally, the operation of enhancing the medical images using our proposed method. The simulation experiments prove that our proposed algorithm achieves superior performance qualitatively and quantitatively, compared with the state-of-the-art experimental methods, Upon a thorough examination and comparative analysis of performance parameters. Furthermore, the advantageous characteristic of this algorithm is its applicability in multiple types of images. Improving the quality of the medical images using our algorithm allows us to attain a superior visual impact on the processed image, and to increase the rate of conformity in the clinical diagnosis. Our proposed model illustrates the structure and forms of relevant details, contained in the medical images. This leads to an increase in overall contrast and enhances visual perception. [ABSTRACT FROM AUTHOR]

Image processing holds an indispensable role in various facets of our daily lives, professional undertakings, and educational pursuits, encompassing a gamut of tasks including image reconstruction, inpainting, super-resolution, colorization, and editing. In recent years, the advent of advanced models rooted in Generative Adversarial Networks (GANs) has showcased remarkable capabilities in the domain of image synthesis, catapulting the direct application of these cutting-edge models to image processing to the forefront of contemporary research. Within this context, GAN inversion, an emerging paradigm, assumes a pivotal role in the landscape of image processing tasks. This paper delves into the realm of image inversion based on the latent space of GAN models. In response to the inherent limitations of current GAN inversion methods, we introduce three innovations. Firstly, we depart from the conventional use of convolutional networks for generator implementation in existing GAN inversion techniques. Our approach employs generators entirely composed of fully connected layers, marking a significant departure from spatial convolutions and information propagation across pixels. Secondly, we leverage the distinct characteristic of generators engaged in conditional independent pixel synthesis. This feature is enhanced by fusing feature maps spanning contiguous strata of a feature pyramid network during the feature extraction process. Lastly, our framework offers a high degree of versatility, extending its applicability beyond image reconstruction to domains like image inpainting, super-resolution, and image colorization. Empirical results, based on the CelebFaces Attribute-HQ (CelebA-HQ) dataset, unequivocally demonstrate that GAN inversion, built on the principle of conditional independent pixel synthesis, yields superior reconstruction outcomes. Furthermore, it proves amenable to a plethora of tasks, including image inpainting, super-resolution, and image colorization. These advances open new vistas in image processing. [ABSTRACT FROM AUTHOR]

Satellite image generation has a wide range of applications. For example, parts of images must be restored in areas obscured by clouds or cloud shadows or areas that must be anonymized. The need to cover a large area with the generated images faces the challenge that separately generated images must maintain the structural and color continuity between the adjacent generated images as well as the actual ones. This study presents a modified architecture of the generative adversarial network (GAN) pix2pix that ensures the integrity of the generated remote sensing images. The pix2pix model comprises a U-Net generator and a PatchGAN discriminator. The generator was modified by expanding the input set with images representing the known parts of ground truth and the respective mask. Data used for the generative model consist of Sentinel-2 (S2) RGB satellite imagery as the target data and OpenStreetMap mapping data as the input. Since forested areas and fields dominate in images, a Kneedle clusterization method was applied to create datasets that better represent the other classes, such as buildings and roads. The original and updated models were trained on different datasets and their results were evaluated using gradient magnitude (GM), Fréchet inception distance (FID), structural similarity index measure (SSIM), and multiscale structural similarity index measure (MS-SSIM) metrics. The models with the updated architecture show improvement in gradient magnitude, SSIM, and MS-SSIM values for all datasets. The average GMs of the junction region and the full image are similar (do not exceed 7%) for the images generated using the modified architecture whereas it is more than 13% higher in the junction area for the images generated using the original architecture. The importance of class balancing is demonstrated by the fact that, for both architectures, models trained on the dataset with a higher ratio of classes representing buildings and roads compared to the models trained on the dataset without clusterization have more than 10% lower FID (162.673 to 190.036 for pix2pix and 173.408 to 195.621 for the modified architecture) and more than 5% higher SSIM (0.3532 to 0.3284 for pix2pix and 0.3575 to 0.3345 for the modified architecture) and MS-SSIM (0.3532 to 0.3284 for pix2pix and 0.3575 to 0.3345 for the modified architecture) values. [ABSTRACT FROM AUTHOR]