We propose a novel design that incorporates a dual-discriminator strategy: one focused on ensuring global image consistency, and the other on preserving local texture accuracy. Experiments conducted on benchmark datasets demonstrate that our approach surpasses prior solutions both quantitatively and in terms of perceptual quality. The proposed architecture is particularly efective at recovering fine textures and sharp edges while maintaining a natural appearance. We investigate the influence of diferent loss function combinations and training strategies on the fidelity-perceptual quality trade-of. Our findings contribute to the growing field of deep learning-based super-resolution, ofering practical insights into architectural refinements that balance computational eficiency with output quality.
The super-resolution (SR) problem implies the task of recovering low-resolution (LR) images to their
high-resolution (HR) counterpart. The technology has been used in a wide spectrum of applications:
medical imaging, satellite photography, and even surveillance systems. An especially important domain
is unmanned aerial vehicle (UAV) navigation, where high-quality visual data directly influences the
accuracy and reliability of flight control and decision-making. Previous research highlights the role of
intelligent visual navigation systems of high accuracy [
In this paper, we address limitations of existing methods by proposing a novel architecture that extends recent generative adversarial network (GAN) based image synthesis methods and incorporates extensions for super-resolution tasks. Our approach attempts to overcome the present limitations through a dual-discriminator architecture that significantly improves the quantitative and qualitative attributes of super-resolution outputs.
Overall, our main contributions: (1) We propose dual-discriminator architecture which simultaneously addresses two diferent aspects of image reconstruction: local features precision and global image coherence. With this design, our model is able to generate high-resolution outputs with both fine-grained texture details and general image consistency. (2) In advance, we propose to conduct an investigation of diferent loss functions to determine which one has the most impact on GAN performance. This analysis provides insightful information on the inherent trade-ofs between conflicting optimization objectives and has practical relevance to architecture design decisions.
The traditional techniques for super-resolution are mostly consist of diferent interpolation techniques
such as bicubic, bilinear, and nearest neighbor algorithms [
Implementation of deep learning has led to significant advancements in the field of image
superresolution [
With improvements in network architectures, Lai et al. [
Generative Adversarial Networks resulted in a huge leap in super-resolution research. Ledig et al. [18] introduced SRGAN, the first GAN-based approach to photo-realistic super-resolution, with perceptual loss functions along with adversarial training. ESRGAN [19] also refined this approach with residual in residual dense blocks and relativistic adversarial loss. Zhang et al. [20] proposed SFTGAN, which incorporated semantic segmentation information to promote structural consistency. Despite these refinements, GAN-based methods still have a tendency to sufer from loss of small details quality and are hard to train.
Recent research has shown that the utilization of more than one discriminator can enhance GAN performance on various tasks. For example, in the field of image-to-image translation, Isola et al. [ 21] employed a PatchGAN discriminator to successfully model local textures. Taking this concept further, Wang et al. [22] introduced multi-scale discriminators for high-resolution image synthesis. In the context of super-resolution, Sajjadi et al. [23] utilized a texture discriminator alongside a standard discriminator, but focused primarily on texture transfer rather than structural coherence.
Our contribution consist in implementation of dual discriminator, designed to capture both local texture coherence and global image consistency. The local discriminator operates on randomly sampled patches, ensuring realistic textures and preserving fine details, while the global discriminator evaluates the overall composition and perceptual quality of the entire image. By combining these complementary perspectives, the approach efectively overcomes the typical trade-of between sharp, realistic textures and coherent global structure.
The choice of loss function plays crucial role in determining model performance. Conventional pixelwise loss functions like L1 and L2 are efective for optimizing PSNR but often generate overly smooth and less detailed outputs. Johnson et al. [24] initiated perceptual loss, which leverages feature activations from pre-trained network to capture semantic information beyond pixel-level accuracy. Wang et al. [25] proposed the LPIPS metric, which has good correlation with human judgment and was employed as an evaluation metric and training objective.
Adversarial losses have evolved from the initial GAN formulation to more stable formulations, such as WGAN [26], WGAN-GP [27], and relativistic GAN [28]. These formulations mitigate training instability often encountered in GAN-based super-resolution. Additionally, various approaches combine diferent content losses, adversarial losses, and regularization terms to achieve an optimal balance between reconstruction fidelity and perceptual quality.
Our proposed method advances these baseline approaches by addressing their limitations through a novel dual-discriminator architecture, tailored specifically for super-resolution task. We conduct a comprehensive analysis of various loss function combinations to identify training strategies that optimally balance reconstruction accuracy with perceptual quality.
The primary objective is to reconstruct low-resolution input image and generate its super-resolved counterpart . During training, the high-resolution serves as groun truth reference. Bicubic interpolation is used to downsample images and create the low-resolution inputs.
Following Goodfellow et al. [29], the GAN framework can be defined as a generator and discriminator . The primary goal is to optimize generator together with discriminator to address the adversarial min-max problem (1).
min max E∼ data ()[log ()] + E∼ ()[log(1 − (()))] (1)
Here () represents the distribution of real data samples, () is the prior distribution of input noise variables, () generates fake samples, and () outputs the probability that came from the real data distribution rather than the generator’s distribution. In the context of super-resolution, the generator maps low-resolution images to high-resolution counterparts, while the discriminator attempts to distinguish between real high-resolution images and the generator’s super-resolved outputs.
SRGAN has a conventional GAN structure: a discriminator and a generator, without auxiliary models, but their structures are modified to address super-resolution task Fig. 1 and 2.
For our research, we will adopt SRGAN as the foundational model. This choice is driven by its proven ability to generate high-resolution images. By leveraging SRGAN’s architecture, which combines adversarial training with perceptual loss, we can ensure that the resulting images not only have improved pixel density but also maintain visual realism. Its established success in super-resolution tasks provides a reliable starting point for further modifications and optimizations tailored to our specific objectives.
Building upon SRGAN’s architecture, the modified model for our task retains the original generator to preserve its high-quality image reconstruction capabilities. The primary change lies in the discriminator, which is replaced by a dual-discriminator setup designed to evaluate diferent aspects of the generated images Fig. 3. This modification enables us to investigate how employing two complementary discriminators influences both the training dynamics and the final image quality.
Generator G takes a low-resolution image as an input and upscales it to produce super-resolution image . Two discriminators 1 and 2, then receive both generated super-resolution image and original high-resolution image as inputs. Their task is to distinguish between real high-resolution and synthetically generated image
The structure of the global discriminator consists of three main parts: a downsampling block, a bottleneck and a residual block. Their interaction allows for efective classification of the input images. The architecture of the global discriminator is illustrated in Fig. 4.
The network takes as input an image that passes through an initial convolutional layer with a kernel size of 4x4 and 64 feature maps, and then a Leaky ReLU layer. Next, let’s take a closer look at the structure of each block and the task they perform when processing the input image.
The first part is the downsampling block. It consists of 5 blocks of a similar structure, including a convolutional layer, batch normalization, Leaky ReLU, but in each subsequent block the number of feature maps doubles. Such structure allows to gradually reduce the image discretization, which allows the network to obtain image features at diferent sizes. Initial blocks receive basic patterns such as edges and textures, when deeper blocks can recognize higher-level semantic features.
Next part of the network is the sifting block. It also consists of a standard structure: a convolutional layer with a kernel dimension of 1x1 and 1024 feature maps, a batch normalization, LeakyReLU, a convolutional layer with a kernel dimension of 1x1 and 512 feature maps, and another batch normalization. Its task is to compress information and create cross-channel integration. It also reduces the number of parameters and thus the cost of calculations.
The last part is the residual block. It contains of a convolutional layer of 1x1 kernels and 128 feature maps, followed by a layer of batch normalization, Leaky ReLU, which are repeated twice with the only diference being the size of the 3x3 kernel of the convolutional layer. After, there is a convolutional layer of 3x3 kernels and 512 feature maps, batch normalization, and the elementwise sum. Its implementation helps prevent the problem of vanishing gradients.
In the final stage of the global discriminator, the output from the preceding blocks is passed through a LeakyReLU activation, followed by a flattening layer and a linear activation function. This produces a ifnal prediction, indicating whether the analyzed input image is classified as real or fake.
Unlike the global discriminator, which evaluates the image as a whole, the local discriminator focuses on analyzing smaller regions. This approach enables the model to examine fine-grained features in greater detail, such as textures, surface patterns, object edges, and other local structures.
By assessing these localized areas, the discriminator can provide the generator with precise feedback on the quality of each region. This, in turn, helps enhance local detail—particularly important when restoring images containing numerous small objects and intricate features that might otherwise be lost during resolution enhancement.
The local discriminator has more compact architecture compared to the global discriminator, consisting primarily of a single feature extraction block. It processes several cropped patches of fixed size, as specified in the network parameters. The overall structure of the local discriminator is illustrated in Fig. 5.
Sliding window with determined parameters picks parts of the images, which are used as input of the local discriminator. This parts of image pass through the following initial layers: convolutional with 4x4 kernel size and 64 feature maps and Leaky ReLU layer. Next, the main part begins, where features are extracted from the input data. It consists of three blocks of similar structure: a convolutional layer with 4x4 kernel size, batch normalization, and Leaky ReLU. The only diference between them is in the size of the feature maps, which, with increasing depth, have the following sizes: 128, 256, and 512. Finally, the obtained features are passed through a straightening layer and a linear activation function.
Similarly to the global discriminator, the 4x4 kernel size allows us to maintain a balance between the complexity of calculations and the capture of local features. This convolutional size also prevents any loss of features.
The combination of global and local discriminators improves the image analysis and thus increases the percentage of correctly identified images. As a result, the generator has much more feedback, and therefore better results.
A sliding window, configured with predefined parameters, extracts image patches that serve as inputs for the local discriminator. These patches first pass through an initial processing stage consisting of a convolutional layer with a 4x4 kernel, 64 feature maps, and a Leaky ReLU activation.
The main feature extraction stage follows, comprising three sequential blocks of identical structure. Each block contains a convolutional layer with a 4x4 kernel, batch normalization, and a Leaky ReLU activation. The primary diference between these blocks lies in the number of feature maps, which increase with network depth: 128, 256, and 512, respectively. The extracted features are then flattened and passed through a fully connected layer with a linear activation function.
By combining the outputs of both the global and local discriminators, the network benefits from additional perspectives: global structure evaluation and fine-grained local analysis. This synergy improves image assessment accuracy, providing the generator with richer and more precise feedback, that leads to higher-quality results.
Defining the loss function is a crucial factor for the quality of the network. In many similar works, a common option is to use the mean square error (MSE) loss function [30, 31]. In this work, we use a modified version of MSE (2), which incorporates additional image analysis through a separate VGG19 network [32].
ℒ = ℒ + ℒ + ℒ (2)
Here ℒ is the common MSE loss function, ℒ is the loss function associated with the VGG19 network and ℒ is a adversarial loss. 2.5.1. MSE loss The pixel-by-pixel MSE loss (3) is well known function which measures the average squared diference between corresponding pixels in predicted () and target images of width and height . ℒ = =1 =1 1 ∑︁ ∑︁ (︁ , − G (︀ )︀ ,︁) 2 (3)
Since this loss function cannot provide enough information to support the recovery of small image details, resulting in smoothing of textures and poor appearance of the generated image, we introduce an additional perceptual loss function based on VGG19 [33], which will bring the results closer to generating realistic images. 2.5.2. Perceptual loss function
ℒ = 2 * 10 −6 · ︀( ⃦⃦ (︀ )︀ − (︀ )︀ ⃦⃦ 2︀) 2 (4)
As previously noted, the perceptual loss function (4) uses a pre-trained VGG19 neural network [32] to extract high-level feature representations of images. It is defined as the Euclidean distance between the feature maps produced by VGG network (·) when applied to real image ( and generated image . 2.5.3. Adversarial loss
Here Global () and Local () are results of the local and global discriminators, respectively and ℒ is Binary Cross-Entropy (BCE) [34].
In addition we also utilize the adversarial part of our GAN network. The associated loss function drives the generator tooward producing more natural-looking images generation capable of deceiving the network’s dual discriminator. Given that our model utilizes two discriminators, the loss function is shown in (5), (6), and (7).
The overall loss function efectively improves high-resolution image generation by integrating three components: MSE, VGG, and adversarial loss. This integration achieves an optimal balance between reconstruction accuracy, fine-detail preservation, and perceptual realism, making it a highly efective approach for generating high-quality images.
The dataset used for training the network is the DOTA dataset [35]. This dataset contains 2806 aerial images collected from various sensors and platforms, each of them is approximately 4000×4000 pixels in size, although the dataset includes images ranging from 800×800 to 20,000×20,000 pixels.
The images are taken from several sources, including Google Earth, GF-2 and JL-1 satellites, and others. The dataset contains both RGB and grayscale images, providing a variety of spectral information for analysis.
DOTA covers a wide range of geographic locations, such as airports, agricultural fields, urban areas, etc. and objects - airplanes, ships, cars, buildings, etc. Wide range of scales, orientations, and shapes enriches training data, reducing overfitting and supporting efective image reconstruction across both ifne and coarse details.
However, original dataset does not fully meet the requirments for network training must be preprocessed. Each image will be cropped to a fixed size of 1024x1024 pixels. Additionally, since some images contain black borders, further actions will be applied to remove these artifacts and avoid potential training issues. Examples of processed images are illustrated in Fig. 6.
After preprocessing, the dataset contains 1,900 images, which are split into training (70%), validation (20%), and test (10%) subsets. This refined dataset supports efective model training and reliable performance evaluation.
Initial training process for the generator lasted 50 epochs, after which adversarial training for double discriminator GAN was conducted witrh a length of 1200 epochs with training parameters: • Batch size: 16. • Learning rate decay: 0.1.
• Optimizer: Adam with parameters ℎ = 10−4 ; 1 = 0.9; 2 = 0.999.
Model evaluated its performance every 5 epochs on a validation subset. According to visual assessment, the model demonstrated quite good results. Fig. 7 demonstrates the results of increasing image resolution.
In this image, the left column shows the original low-resolution images from the dataset, the center column presents the results produced by the modified generative adversarial network, and the right column displays the original high-resolution images.
To evaluate objective quality of super-resolution results, we computed numerical metrics including Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index Measure (SSIM). Table 1 compares these values between initial and final training epochs.
The results confirm the efectiveness of the developed modified generative adversarial network for satellite image super-resolution. Similar metric values between training and test sets indicate good model generalization without overfitting. Despite limitations in recovering fine details in complex images, the model shows stable performance and is suitable for practical satellite image enhancement with 4x resolution upscaling.
To further assess efectiveness of the dual discriminator model, we conducted an additional comparison with a baseline version of the network containing only global discriminator. For a fair evaluation, both models were configured identically and trained for the same number of epochs. After training the baseline model for 1200 epochs, the results presented in Table 3 were obtained.
The results clearly indicate that proposed double discriminator model outperforms base model in all metrics. Moreover, the diference in the obtained values is significant.
To evaluate the performance of our dual discriminator approach against state-of-the-art super-resolution methods, we conducted a comparative analysis with RealESRGAN, a widely recognized model for realworld image super-resolution. RealESRGAN employs a U-Net generator with skip connections and spectral normalization, trained with a combination of L1 loss, perceptual loss, and adversarial loss using a single discriminator architecture.
For fair comparison, both models were evaluated on the same test dataset containing satellite images with 4x upscaling factor. The RealESRGAN model was fine-tuned on our satellite image dataset to ensure optimal performance for this specific task.
The quantitative results demonstrate that our dual discriminator architecture achieves superior performance across all evaluated metrics. Our model consistently outperforms RealESRGAN across all metrics.
Visual comparison reveals that our approach produces sharper edges and better preserves finegrained details in satellite imagery, particularly in areas with complex textures such as urban regions and vegetation boundaries. RealESRGAN produces smoothed results that lose important structural information crucial for satellite image analysis applications. Fig. 8 presents a qualitative comparison between the two methods.
The improved performance can be attributed to the complementary roles of global and local discriminators in our architecture. While the global discriminator ensures overall image coherence similar to RealESRGAN’s approach, the local discriminator specifically focuses on enhancing fine details and texture quality, preventing smoothing and leading to more realistic and detailed super-resolution results.
We propose a novel approach for enhancing the resolution of satellite images using a GAN architecture with dual discriminators. Each discriminator targets diferent quality features, addressing existing limitations in image enhancement and enabling more precise visual improvements.
The dual discriminator approach improves training stability and mitigates mode collapse, while the use of a diverse training dataset increases model’s adaptability in diferent satellite conditions, such as changing lighting conditions, seasonal fluctuations, and geographical terrain.
Experimental evaluations on both validation and test datasets show strong performance. For the test set, the model achieved MSE = 0.0089, PSNR = 26.95 dB, and SSIM = 0.7663. The generated images demonstrate high visual fidelity and photorealistic detail.
This research holds significant potential for practical applications in environmental monitoring, urban planning, emergency response, and crop assessment and other fields where high-resolution imagery is important for decision-making.
In conclusion, the proposed dual-discriminator GAN ofers an efective and stable solution for satellite image super-resolution, combining computational eficiency with substantial improvements in image quality.
During the preparation of this work, the author(s) used GPT-4o in order to: Grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2472-2481). [18] Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., ... & Shi, W. (2017). Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4681-4690). [19] Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., ... & Change Loy, C. (2018). Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops (pp. 0-0). [20] Zhang, Y., Li, X., & Zhou, J. (2019). SFTGAN: a generative adversarial network for pan-sharpening equipped with spatial feature transform layers. Journal of Applied Remote Sensing, 13(2), 026507026507. [21] Isola, P., Zhu, J. Y., Zhou, T., & Efros, A. A. (2017). Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1125-1134). [22] Wang, T. C., Liu, M. Y., Zhu, J. Y., Tao, A., Kautz, J., & Catanzaro, B. (2018). High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 8798-8807). [23] Sajjadi, M. S., Scholkopf, B., & Hirsch, M. (2017). Enhancenet: Single image super-resolution through automated texture synthesis. In Proceedings of the IEEE international conference on computer vision (pp. 4491-4500). [24] Johnson, J., Alahi, A., & Fei-Fei, L. (2016). Perceptual losses for real-time style transfer and super-resolution. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14 (pp. 694-711). Springer International Publishing. [25] Wang, Z., Simoncelli, E. P., & Bovik, A. C. (2003, November). Multiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003 (Vol. 2, pp. 1398-1402). Ieee. [26] Arjovsky, M., Chintala, S., & Bottou, L. (2017, July). Wasserstein generative adversarial networks.
In International conference on machine learning (pp. 214-223). PMLR. [27] Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., & Courville, A. C. (2017). Improved training of wasserstein gans. Advances in neural information processing systems, 30. [28] Jolicoeur-Martineau, A. (2018). The relativistic discriminator: a key element missing from standard
GAN. arXiv preprint arXiv:1807.00734. [29] Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., ... & Bengio, Y.
(2014). Generative adversarial nets. Advances in neural information processing systems, 27. [30] Dong, C., Loy, C. C., He, K., & Tang, X. (2015). Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2), 295-307. [31] Shi, W., Caballero, J., Huszár, F., Totz, J., Aitken, A. P., Bishop, R., ... & Wang, Z. (2016). Real-time single image and video super-resolution using an eficient sub-pixel convolutional neural network.
In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1874-1883). [32] Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556. [33] Gatys, L., Ecker, A. S., & Bethge, M. (2015). Texture synthesis using convolutional neural networks.
Advances in neural information processing systems, 28. [34] Ruby, U., & Yendapalli, V. (2020). Binary cross entropy with deep learning technique for image classification. Int. J. Adv. Trends Comput. Sci. Eng, 9(10). [35] Xia, G. S., Bai, X., Ding, J., Zhu, Z., Belongie, S., Luo, J., ... & Zhang, L. (2018). DOTA: A large-scale dataset for object detection in aerial images. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 3974-3983).