High Visual Quality Research Articles

Penyisipan Data Secara Reversible Pada Gambar Medis Menggunakan Metode Difference Expansion of Quad dan Fungsi Modulus

The rapid advancement of the Internet of Things (IoT) brings convenience across various domains, but raises its own challenges, particularly concerning data security and privacy. Cryptography and steganography are two techniques used to secure data. Cryptography protects information by encrypting messages, while steganography hides the existence of data in other media. The embedding of data into other media like medical images necessitates specialized approaches to prevent potential distortion that may compromise their integrity. Distortion in medical images could lead to erroneous diagnoses. In this study, we introduce a reversible data embedding technique capable of returning storage media to its original state without distortion, utilizing the Difference Expansion of Quad and Modulus Function. This approach employs 2×2 pixel blocks to embed 3-bits of data. Our experiments on six medical images demonstrate that the proposed method offers substantial embedding capacity while maintaining high visual quality. The embedding capacity can reach 0.4834 bits per pixel (bpp) with a PSNR of 49.4330 dB. Experimental findings indicate that the proposed method outperforms previous techniques.

StegFormer: Rebuilding the Glory of Autoencoder-Based Steganography

Image hiding aims to conceal one or more secret images within a cover image of the same resolution. Due to strict capacity requirements, image hiding is commonly called large-capacity steganography. In this paper, we propose StegFormer, a novel autoencoder-based image-hiding model. StegFormer can conceal one or multiple secret images within a cover image of the same resolution while preserving the high visual quality of the stego image. In addition, to mitigate the limitations of current steganographic models in real-world scenarios, we propose a normalizing training strategy and a restrict loss to improve the reliability of the steganographic models under realistic conditions. Furthermore, we propose an efficient steganographic capacity expansion method to increase the capacity of steganography and enhance the efficiency of secret communication. Through this approach, we can increase the relative payload of StegFormer to 96 bits per pixel without any training strategy modifications. Experiments demonstrate that our StegFormer outperforms existing state-of-the-art (SOTA) models. In the case of single-image steganography, there is an improvement of more than 3 dB and 5 dB in PSNR for secret/recovery image pairs and cover/stego image pairs.

Advanced Dual Reversible Data Hiding: A Focus on Modification Direction and Enhanced Least Significant Bit (LSB) Approaches

In this study, we investigate advances in reversible data hiding (RDH), a critical area in the era of widespread digital data sharing. Recognizing the inherent vulnerabilities such as unauthorized access and data corruption during data transmission, we introduce an innovative dual approach to RDH. We use the EMD (Exploiting Modification Direction) method along with an optimized LSB (Least Significant Bit) replacement strategy. This dual method, applied to grayscale images, has been carefully developed to improve data hiding by focusing on modifying pixel pairs. Our approach sets new standards for achieving a balance between high data embedding rates and the integrity of visual quality. The EMD method ensures that each secret digit in a 5-ary notational system is hidden by 2 cover pixels. Meanwhile, our LSB strategy finely adjusts the pixels selected by EMD to minimize data errors. Despite its simplicity, this approach has been proven to outperform existing technologies. It offers a high embedding rate (ER) while maintaining the high visual quality of the stego images. Moreover, it significantly improves data hiding capacity. This enables the full recovery of the original image without increasing file size or adding unnecessary data, marking a significant breakthrough in data security.

PADVG: A Simple Baseline of Active Protection for Audio-Driven Video Generation

Over the past few years, deep generative models have significantly evolved, enabling the synthesis of realistic content and also bringing security concerns of illegal misuse. Therefore, active protection for generative models has been proposed recently, aiming to generate samples with hidden messages for future identification while preserving the original generating performance. However, existing active protection methods are specifically designed for generative adversarial networks (GANs), restricted to handling unconditional image generation. We observe that they get limited identification performance and visual quality when handling audio-driven video generation conditioned on target audio and source input to drive video generation with consistent context, e.g., identity and movement, between frame sequences. To address this issue, we introduce a simple yet effective active P rotection framework for A udio- D riven V ideo G eneration, named PADVG. To be specific, we present a novel frame-shared embedding module in which messages to hide are first transformed into frame-shared message coefficients. Then, these coefficients are assembled with the intermediate feature maps of video generators at multiple feature levels to generate the embedded video frames. Besides, PADVG further considers two visual consistent losses: (i) intra-frame loss is utilized to keep the visual consistency with different hidden messages; (ii) inter-frame loss is used to preserve the visual consistency across different video frames. Moreover, we also propose an auxiliary denoising training strategy through perturbing the assembled features by learnable pixel-level noise to improve identification performance, while enhancing robustness against real-world disturbances. Extensive experiments demonstrate that our proposed PADVG for audio-driven video generation can effectively identify the generated videos and achieve high visual quality.

Radiation reduction for interventional radiology imaging: a video frame interpolation solution

PurposeThe aim of this study was to diminish radiation exposure in interventional radiology (IR) imaging while maintaining image quality. This was achieved by decreasing the acquisition frame rate and employing a deep neural network to interpolate the reduced frames.MethodsThis retrospective study involved the analysis of 1634 IR sequences from 167 pediatric patients (March 2014 to January 2022). The dataset underwent a random split into training and validation subsets (at a 9:1 ratio) for model training and evaluation. Our approach proficiently synthesized absent frames in simulated low-frame-rate sequences by excluding intermediate frames from the validation subset. Accuracy assessments encompassed both objective experiments and subjective evaluations conducted by nine radiologists.ResultsThe deep learning model adeptly interpolated the eliminated frames within IR sequences, demonstrating encouraging peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) results. The average PSNR values for angiographic, subtraction, and fluoroscopic modes were 44.94 dB, 34.84 dB, and 33.82 dB, respectively, while the corresponding SSIM values were 0.9840, 0.9194, and 0.7752. Subjective experiments conducted with experienced interventional radiologists revealed minimal discernible differences between interpolated and authentic sequences.ConclusionOur method, which interpolates low-frame-rate IR sequences, has shown the capability to produce high-quality IR images. Additionally, the model exhibits potential for reducing the frame rate during IR image acquisition, consequently mitigating radiation exposure.Critical relevance statementThis study presents a critical advancement in clinical radiology by demonstrating the effectiveness of a deep neural network in reducing radiation exposure during pediatric interventional radiology while maintaining image quality, offering a potential solution to enhance patient safety.Key points• Reducing radiation: cutting IR image to reduce radiation.• Accurate frame interpolation: our model effectively interpolates missing frames.• High visual quality in terms of PSNR and SSIM, making IR procedures safer without sacrificing quality.Graphical

On-the-fly image-level oversampling for imbalanced datasets of manufacturing defects

Visual defect recognition and its manufacturing applications have been an upcoming topic in recent AI research. Defect datasets are often severely imbalanced and can be additionally burdened with separating classes of high visual similarity. Although various methods of data augmentation have been proposed to mitigate the class imbalance, they often fail to cope with tinier minority classes or have fidelity issues with smaller defects while, at the same time, needing significant computational resources to train. Also, augmentation based on vector-based oversampling struggles to produce high-fidelity inputs and is hard to apply on custom CNN architectures, which often perform better for this type of problem. Our work presents an image-level oversampling method based on an instance-based image generator that can be applied to any CNN directly during the training process without increasing the order of training time required. It is based on identifying a small number of the most uncertain base samples close to the estimated class boundaries and using them as seeds for augmentation. The resulting images are of high visual quality preserving small class differences, and they also improve the classifier boundary leading to higher recall scores than other state-of-the-art approaches.

A High-Payload Data Hiding Scheme Based on Absolute Moment Block Truncation Coding for Minimizing Hiding Impact

Data hiding encompasses a wide range of applications related to hiding messages in digital images. The visual redundancy of images makes it possible to embed data in the images without attracting attention. Increasing the hiding capacity and decreasing the hiding distortion are prime objectives that data hiding intends to achieve. In this paper, we propose a high-payload data hiding scheme based on absolute moment block truncation coding (AMBTC) to minimize the impact of hiding. A two-level minimum mean square error (MMSE) quantizer generated by AMBTC is used to decrease the distortion associated with hiding. Also, we present a lookup table based on the symmetric property for adaptively hiding secrets in pixels to achieve high hiding capacity. We can embed almost 1.9 bits per pixel (bpp) with a high image quality of an average of 31 dB. Only 5.3% of pixels are changed during the data-hiding process. Compared with other schemes, we can use 1 bpp more relative payload for embedding with the same stego image quality. The experimental results show that the proposed scheme has better hiding performance because it allows a huge amount of secret data to be hidden while maintaining the high visual quality of the stego image.

Verifiable visually meaningful image encryption algorithm based on compressive sensing and (t, n)-threshold secret sharing

Abstract A verifiable visually meaningful image encryption algorithm based on compressive sensing and (t, n)-threshold secret sharing is proposed. Firstly, the plain image is compressed and encrypted by 2D block compressive sensing to obtain the pre-encrypted image. During this process, the enhanced logistic map and the enhanced tent map are used to generate the measurement matrix and the permutation sequence, respectively. Secondly, multiple shadow images of the pre-encrypted image are generated by using the (t, n)-threshold secret sharing scheme, and then each shadow image is further encrypted by using the encryption matrix generated by the chaotic system. Meanwhile, the signatures of the shadow images are obtained by using RSA signature algorithm. Finally, the shadow images and their corresponding signatures are embedded into the carrier image to generate the cipher image with high visual quality by using the LSB method. Additionally, the present algorithm can resist known-plaintext and chosen-plaintext attacks by incorporating the hash value of the plain image as part of the key. Meanwhile, the use of 2D block compressive sensing significantly reduces the reconstruction time. Simulation results demonstrate that the proposed algorithm achieves excellent decryption quality and operational efficiency.

Reconstructing higher-resolution four-dimensional time-varying volumetric data

We have witnessed substantial growth in super-resolution research within the computer vision community. Unlike previous works that mainly focus on the super-resolution synthesis of images, videos, or single volumes, our research is dedicated to the super-resolution synthesis of time-varying volumetric data, which are generated from scientific simulations and are crucial for domain scientists to understand and analyse complex scientific phenomena. Compared to previous works, our research presents a greater challenge: the time-varying volumetric data have higher dimensions, making it more difficult to synthesise super-resolution that maintains good spatio-temporal consistency while achieving high visual quality. To tackle this challenge, we introduce a new GAN-based network called SSR-DoubleUNetGAN, which includes novel network architecture and loss functions, allowing for accurate synthesis of spatial super-resolution for time-varying volumetric data with relatively fast training time. Our method can be applied in the context of in-situ visualisation to aid domain scientists in analysing more time-varying volumetric data more efficiently. In addition, it can be used in the compression-decompression pipeline to recover the super-resolution time-varying volumetric data from their low-resolution counterpart. To demonstrate its effectiveness, we applied various time-varying volumetric datasets from different scientific simulations to it. To demonstrate its advantages, we compared it qualitatively and quantitatively with five state-of-the-art super-resolution techniques, namely SSR-TVD, Tricubic, SRResNet, Cubic, and Linear. Furthermore, we conducted an ablation study to validate its important modules. The experimental results show that our method outperforms the compared state-of-the-art techniques.

Optimized steganography techniques based on PVDS and genetic algorithm

Due to the rapid development of confidential digital communication over international networks, security is the most significant factor. Steganography is the science of concealed communication and is used to conceal the presence of secret information among the various methods of hiding data in communications. Steganography techniques are usually implemented either in spatial or frequency domains. Some spatial domain methods fundamentally utilize the absolute difference value between neighbor pixels to encode the secret message, categorized as Pixel-Value Differencing Steganography methods (PVDS). In PVDS, the hiding capacity of the adjacent pixels depends on their difference value. Consequently, the secret data will be embedded with a high visual quality of the stego-image. Indicator-based PVDS (IPVDS) technique is derived from the PVDS method. It depends on an Indicator pixel (IP) to extract the secret data correctly. The Genetic Algorithm (GA) approach to optimization depends on the concept of survival of the fittest. This paper introduces two proposals, GA-based steganography schemes (GA-IPVD) and (GA-IPVDM). These new techniques encoded secret data using the IPVDS method. Before embedding data in the cover image, the order of the pixels corresponding to the secret data is modified and rearranged. GA controls the operation of rearranging and modifying the order of pixels and all parameters. The current paper addresses steganography as an optimization problem and seeks to determine the optimal order of the pixels that enhance the matching between the cover image and stego-image to reduce distortion despite high embedded capacity. In the first proposed technique GA-IPVD, GA concentrates on the imperceptibility of embedded data and maximizes the quality of the stego-image according to the given secret data. In contrast, the GA maximizes the capacity of embedded data while maintaining an acceptable amount of stego-image quality in the second proposed technique GA-IPVDM. The proposed systems have embedding key that increase the security of proposed schemes. The experimental results demonstrate that the proposed methods can embed a large quantity of secret data while maintaining an acceptable visual quality of the stego-images. The proposed technique has also demonstrated excellent resistance to a variety of stego-attacks, including the pixel difference histogram (PDH) as well as regular and singular (RS) analyses. Additionally, the out-of-boundary pixel issue (BI), which endures in the majority of modern data hiding algorithms, has been successfully resolved.

Comparison of visual performance with iTrace analyzer following femtosecond laser-assisted cataract surgery with bilateral implantation of two different trifocal intraocular lenses.

To compare the postoperative binocular visual performance with an iTrace analyzer following femtosecond laser-assisted cataract surgery (FLACS) combined with bilateral implantation of two different types of diffractive trifocal intraocular lenses (IOL). During this retrospective observational study, patients who received bilateral FLACS combined with implantation of two different types of diffractive trifocal IOLs were evaluated. According to the IOLs' different types and design, the patients were divided into AT LISA tri839MP group (tri839 group) and AcrySof PanOptix TFNT00 group (TFNT group). Study parameters included preoperative and postoperative uncorrected distance visual acuity (UDVA) at 5 m, uncorrected near visual acuity (UNVA) at 30 cm and 40 cm, uncorrected intermediate visual acuity (UIVA) at 60 cm and 80 cm, postoperative refractive status, objective visual qualities and total high order aberrations (HOAs) postoperatively. The postoperative complications were also recorded. Totally 56 eyes of 28 patients (tri839 group, n=26; TFNT group, n=30) were included. Preoperative baseline characteristics between groups were not statistically significantly different. UDVA was not significantly different between groups except for 1wk follow-up due to the postoperative corneal edema. TFNT group showed statistically significant better UNIA at 60 cm than tri839 group at the 1wk (0.05±0.19 vs 0.15±0.10 logMAR, P=0.013), 1mo (0.05±0.12 vs 0.15±0.09 logMAR, P=0.001) and 3mo (0.04±0.12 vs 0.15±0.11 logMAR, P=0.001) follow-up, while tri839 group showed statistically significant better UNIA at 80 cm than TFNT group at the 1d (0.14±0.15 vs 0.20±0.14 logMAR, P=0.041) and 1mo (0.09±0.07 vs 0.14±0.10 logMAR, P=0.042) follow-up. Postoperative refractive status showed stable at every visit. Modulated transfer function (MTF) values and strehl ratio (SR) values were improved and HOAs were lower significantly after surgery. FLACS with bilateral implantations of both tri839 and TFNT00 can achieve satisfactory natural whole-course vision, high postoperative refractive stability and good visual quality but without significantly difference. iTrace aberration instrument can accurately evaluate the visual quality under different status.

Cognitive-emotional feasibility of the effect of visual quality of building form on promoting the sense of place attachment (Case study: Cultural iconic buildings of Iran's contemporary architecture)

The present research studies the relationship between place attachment and the perception of form's visual quality in fifteen outstanding contemporary Iranian architectural cultural buildings. This study puts forward the hypothesis that there is a correlation between the quality of building form and the sense of place attachment, in the sense that creating high visual quality through enhancing the quality of building form increases citizens' initial satisfaction with and subsequent attachment to the building. High visual quality influences people's experience of the environment and improves the quality of life. Place attachment highlights how people, on a personal level, recreate a sense of place for themselves. The present study adopts the descriptive-analytical method as its theoretical framework and the survey as the empirical methodology. Questionnaires were developed using the Likert scale and distributed among experts and ordinary citizens. Data analysis using SPSS and the adoption of descriptive-analytical statistics, correlation analysis, and regression showed the relationship among the characteristics of indicators. The results show a positive correlation between form and place attachment mediated through visual quality, and they are causal conditions for one another. In addition, only some of the buildings under study evoke the same level of place attachment.

A depth iterative illumination estimation network for low-light image enhancement based on retinex theory

Existing low-light image enhancement techniques face challenges in achieving high visual quality and computational efficiency, as well as in effectively removing noise and adjusting illumination in extremely dark scenes. To address these problems, in this paper, we propose an illumination enhancement network based on Retinex theory for fast and accurate brightening of images in low-illumination scenes. Two learning-based networks are carefully constructed: decomposition network and enhancement network. The decomposition network is responsible for decomposing the low-light input image into the initial reflectance and illumination map. The enhanced network includes two sub-modules: the illumination enhancement module and the reflection denoising module, which are used for efficient brightness enhancement and accurate reflectance. Specially, we have established a cascaded iterative lighting learning process and utilized weight sharing to conduct accurate illumination estimation. Additionally, unsupervised training losses are defined to improve the generalization ability of the model. The proposed illumination enhancement framework enables noise suppression and detail preservation of the final decomposition results. To establish the efficacy and superiority of the model, on the widely applicable LOL dataset, our approach achieves a significant 9.16% increase in PSNR compared to the classical Retinex-Net, and a remarkable enhancement of 19.26% compared to the latest SCI method.

Evaluasi Kualitas Visual Lanskap di Pantai Jimbaran, Kabupaten Badung, Bali

Evaluation of Visual Quality of Tourist Landscape in Jimbaran Beach, Badung Regency. Jimbaran Beachas one of the tourist attractions in Badung Regency, Bali does not yet have an optimal arrangement of theconstituent elements of the landscape. This affects the visual comfort of its users. So it is important to improvethe visual beauty of jimbaran beach tourist attraction. Based on this, this study aims to assess the visual qualityof the Jimbaran beach landscape and provide recommendations for structuring the constituent elements of thelandscape. The Method used in this research is mix-method which are qualitative by analyzing based on theexisting situation and standard from related teories and quantitative to find out the visual quality of landscapeelements by a method called Scenic Beauty Estimation (SBE). The results of the overall assessment byrespondents showed that the visual quality of Jimbaran beach was included in the moderate category.Recommendations for landscape arrangements are needed based on the high visual quality of jimbaran beachlandscape. The recommendation is expected to be an input for managers to improve the visual quality of theJimbaran beach landscape.

Comparing the effects of immersive and non-immersive real estate experience on behavioral intentions

Nowadays, the web presents real estate sites providing the possibility to use low-cost virtual reality (VR) solutions without sacrificing the high quality of graphic visualization of their contents to attract customers through enhanced digital experiences and increase the probability of selling their products. However, whether immersive VR technologies improve the chances of selling real estate is unclear. This study examines the effects induced by a real estate website that offers the opportunity to evaluate the potential purchase of a property in two ways: by viewing photos of the house (non-immersive conditions) and by taking a virtual tour of the same house using a VR head-mounted display (immersive conditions). Forty participants interested in evaluating the purchase of a new house, randomly assigned to one of the experimental conditions, participated in the experiment. During the real estate experience, electroencephalography (EEG) and skin conductance (SC) were recorded to collect biological data about the emotional effects raised by each condition. After the study, psychological self-reported questionnaires were collected measuring: emotions, presence, user experience (UX), and a three-item ad hoc scale exploring behavioral intentions (e.g., the perceived pleasantness of the house, the willingness to visit the house, and the purchase intention). Results show that immersive navigation is associated with more positive emotions (EEG), more intense physiological arousal (SC), higher scores on the presence questionnaire, and higher scores on the UX questionnaire compared to the non-immersive one. The psychological scales measuring emotions did not yield significant results. No significant difference in the three-item ad-hoc self-questionnaire was found, except for the item examining willingness to visit the real house building. Implications are also discussed.

Reinforcement learning for high-quality reality mapping of indoor construction using unmanned ground vehicles

Recent advances in reality capture technology focused on automating reality capture and devising robust computational models to convert the collected data into usable formats. However, these modern approaches are still challenged by the insufficient capacity of the resulting information to relay accurate and complete representations of the construction state due to the data's poor visual quality. In addition, the complexity of robotic path planning –especially for indoor construction– hinders automatic visual data collection due to cluttered, narrow, and dynamic construction spaces. This work targets both challenges by presenting a reinforcement learning model that optimizes indoor data collection policy for acquiring high-quality visual data using camera-equipped unmanned ground rovers. Results from three learned navigation policies show the capability of the method to provide high visual quality for the collected data. The learned policies reduced the data collection duration by 38.23% on average compared to the currently used automatic data collection strategies. The policies also provided a 31.04% average reduction in data collection distance compared to lawn-mower patterns.

An Improved Image Compression Algorithm Using 2D DWT and PCA with Canonical Huffman Encoding.

Of late, image compression has become crucial due to the rising need for faster encoding and decoding. To achieve this objective, the present study proposes the use of canonical Huffman coding (CHC) as an entropy coder, which entails a lower decoding time compared to binary Huffman codes. For image compression, discrete wavelet transform (DWT) and CHC with principal component analysis (PCA) were combined. The lossy method was introduced by using PCA, followed by DWT and CHC to enhance compression efficiency. By using DWT and CHC instead of PCA alone, the reconstructed images have a better peak signal-to-noise ratio (PSNR). In this study, we also developed a hybrid compression model combining the advantages of DWT, CHC and PCA. With the increasing use of image data, better image compression techniques are necessary for the efficient use of storage space. The proposed technique achieved up to 60% compression while maintaining high visual quality. This method also outperformed the currently available techniques in terms of both PSNR (in dB) and bit-per-pixel (bpp) scores. This approach was tested on various color images, including Peppers 512 × 512 × 3 and Couple 256 × 256 × 3, showing improvements by 17 dB and 22 dB, respectively, while reducing the bpp by 0.56 and 0.10, respectively. For grayscale images as well, i.e., Lena 512 × 512 and Boat 256 × 256, the proposed method showed improvements by 5 dB and 8 dB, respectively, with a decrease of 0.02 bpp in both cases.

Joint network for specular highlight detection and adversarial generation of specular-free images trained with polarimetric data

Specular highlights in images pose a significant challenge in algorithms for image segmentation, object detection and other image-based decision-making systems. However, most systems ignore this particular scenario and neglect input images with specular highlights instead of mitigating it in the pre-processing stage. In this paper, we leverage deep neural networks and take advantage of the varying illumination information in polarimetric images for synthesizing specular-free images. We propose a multi-domain Specular Highlight Mitigation Generative Adversarial Network (SHMGAN) with self-attention. SHMGAN consists of a single generator–discriminator pair trained simultaneously using polarimetric images. The proposed GAN uses a dynamically generated attention mask based on a specularity segmentation network without requiring additional manual input. The network is able to learn the illumination variation between the four polarimetric images and a pseudo-diffuse image. Once trained, SHMGAN is able to generate specular-free images from a single RGB image as input; without requiring any additional external labels. The proposed network is trained and tested publicly available datasets of real-world images. SHMGAN is able to accurately identify the specularity affected pixels and generates high visual quality images with mitigated specular reflections. The generated images are realistic and have very low noise, distortions and aberrations compared to the existing state-of-the-art methods for specular highlight removal.

Research on Image Super-Resolution Using Attention Mechanisms based on Super-Resolution Generative Adversarial Network

With the continuous advancement of technology, Super-Resolution Generative Adversarial Networks (SRGAN) have played a significant role in the field of image super-resolution, significantly enhancing the resolution of images. However, while SRGAN excels in generating details, sometimes the restored details do not always meet people's expectations. To further enhance the quality of images and make image details clearer, this paper introduces improvements to the architecture and loss functions of the SRGAN network. Specifically, this research draws inspiration from the architecture of ESRGAN, removing the original Batch Normalization layers and introducing a newly designed Residual Block. Leveraging insights from attention mechanisms, we incorporate three layers of convolutional operations and introduce attention mechanisms into these new Residual Blocks. Furthermore, to simplify the computational complexity of the model, this paper simplifies the original loss functions, consolidating the previous four losses into two. These enhancements result in a significantly improved model in capturing visual elements, making key objects in the images more prominent compared to SRGAN. Detailed experimental results demonstrate that this model, while maintaining the clarity of details, provides higher visual quality. These achievements provide valuable insights and inspiration for further research and applications in the field of image super-resolution.

DeformSg2im: Scene graph based multi-instance image generation with a deformable geometric layout

Recently, scene graph based image generation has emerged to be an important research direction for advanced multi-instance generation tasks. Scene layout generation, the phase that produces the instance-wise visual representations and maintains the spatial relationships among all instances, is vital in translating a scene graph to an image. However, the scene layouts generated by existing methods are too coarse and semantically inconsistent with the given scene graph, which results in quality degradation of the generated images. Motivated by this, we propose a novel scene graph based image generation model called DeformSg2im, which aims to generate visually appealing and semantically faithful images according to the given scene graphs. Our method addresses the aforementioned problems from two aspects. In one aspect, we present an attention-based instance embedding estimator to refine the shape information for each instance. By introducing attention maps to the estimation of instance embeddings, our method is able to generate more plausible instances with sharp edges on the images. In the other aspect, a spatial warping network (SWN) is proposed to adaptively capture the spatial dependencies among instances. With sequential modeling and geometric deformations, the SWN is capable of generating the scene layout conforming to the given scene graph. Extensive experiments show that our model generates images with high visual quality and achieves competitive quantitative results compared to existing works. Ablation studies on the proposed modules are also conducted, and the results demonstrate the effectiveness of our method.