L1 Loss Research Articles

DifFace: Blind Face Restoration With Diffused Error Contraction.

While deep learning-based methods for blind face restoration have achieved unprecedented success, they still suffer from two major limitations. First, most of them deteriorate when facing complex degradations out of their training data. Second, these methods require multiple constraints, e.g., fidelity, perceptual, and adversarial losses, which require laborious hyper-parameter tuning to stabilize and balance their influences. In this work, we propose a novel method named DifFace that is capable of coping with unseen and complex degradations more gracefully without complicated loss designs. The key of our method is to establish a posterior distribution from the observed low-quality (LQ) image to its high-quality (HQ) counterpart. In particular, we design a transition distribution from the LQ image to the intermediate state of a pre-trained diffusion model and then gradually transmit from this intermediate state to the HQ target by recursively applying a pre-trained diffusion model. The transition distribution only relies on a restoration backbone that is trained with L1 loss on some synthetic data, which favorably avoids the cumbersome training process in existing methods. Moreover, the transition distribution can contract the error of the restoration backbone and thus makes our method more robust to unknown degradations. Comprehensive experiments show that DifFace is superior to current state-of-the-art methods, especially in cases with severe degradations.

Removing random noise and improving the resolution of seismic data using deep‐learning transformers

AbstractPost‐stack data are susceptible to noise interference and have low resolution, which impacts the accuracy and efficiency of subsequent seismic data interpretation. To address this issue, we propose a deep learning approach called Seis‐SUnet, which achieves simultaneous random noise suppression and super‐resolution reconstruction of seismic data. First, the Conv‐Swin‐Block is designed to utilize ordinary convolution and Swin transformer to capture the long‐distance dependencies in the spatial location of seismic data, enabling the network to comprehensively comprehend the overall structure of seismic data. Second, to address the problem of weakening the effective signal during network mapping, we use a hybrid training strategy of L1 loss, edge loss and multi‐scale structural similarity loss. The edge loss function directs the network training to focus more on the high‐frequency information at the edges of seismic data by amplifying the weight. Additionally, the verification of synthetic and field seismic datasets confirms that Seis‐SUnet can effectively improve the signal‐to‐noise ratio and resolution of seismic data. By comparing it with traditional methods and two deep learning reconstruction methods, experimental results demonstrate that Seis‐SUnet excels in removing random noise, preserving the continuity of rock layers and maintaining faults as well as being strong robustness.

YOLOrot2.0: A novel algorithm for high-precision rice seed size measurement with real-time processing

We present YOLOrot2.0, an automated algorithm for measuring the dimensions of rice seeds, which holds significant importance in evaluating grain quality and genetic breeding research. The current methods for rice seed measurement exhibit various limitations, such as the need for seed preprocessing and selection, specialized equipment for measurement, and prolonged measurement time. To address these issues, YOLOrot2.0 introduces the following improvements: Firstly, it utilizes an anchor-free detection algorithm to optimize the detection of rotated targets. Secondly, the utilization of the Kalman Filter IoU loss function, which combines the horizontal bounding box and smooth L1 loss function, accelerates the convergence speed of the network. Moreover, the YOLOv8 architecture was adjusted by modifying certain convolutional layers and the Context to Fusion module to enhance the detection capabilities for smaller targets. Finally, YOLOrot2.0 enables direct processing of the entire image, eliminating the need for image partitioning and resulting in substantial time and processing savings. Experimental results demonstrate that YOLOrot2.0 achieves an mAP (mean average precision) score of 0.95, surpassing other comparative algorithms. Additionally, YOLOrot2.0 provides highly accurate measurements of seed length and width, closely aligned with ground truth values. Notably, YOLOrot2.0 achieves real-time detection by achieving an average detection time of only 11.5 ms for images sized 3000×4000 pixels.

Learnable dual attention fusion network for borehole distributed acoustic sensing systems data reconstruction

As the demand for nonrenewable oil and gas resources has increased, distributed acoustic sensing systems (DAS) have been widely used in the acquisition of vertical seismic profiles. Complicated subsurface conditions produce many types of noise with strong energy. The effective reflection and refraction signals are absorbed by the inelastic medium during the transmission process, resulting in significant energy loss and aliased frequency characteristics. However, satisfactory processing results and quality cannot be achieved through conventional denoising methods. In recent years, the rapid development of deep-learning methods with attention mechanisms has attracted significant attention. Noise reduction and signal reconstruction technologies with multiple attention mechanisms within the framework of joint-domain features have become important development directions in DAS data processing. An attention mechanism can help the network ignore useless information and pay attention to useful information in the same way that humans can. In this study, we discuss a novel learnable dual attention fusion network that combines local and patch-based nonlocal attention mechanisms based on ResBlock modules, making the extraction process of the network learnable and selective. In addition, the novel network can focus on local and global details. We also use a residual learning strategy to simplify the training process and gradually establish the best mapping based on the L2 loss function. All experimental results of the simulation and actual records show that our novel network has better denoising and effective signal restoration abilities. Moreover, it can recover the up-going and down-going wavefields and attenuate different types of noise with less effective energy loss.

Ranking-based adaptive query generation for DETRs in crowded pedestrian detection

Variants of DEtection TRansformer (DETRs) have shown promising performance in crowded pedestrian detection. However, we observe that DETRs are sensitive to the hyper-parameter (the number of queries). Adjusting this hyper-parameter is crucial for achieving competitive performance across different crowded pedestrian datasets. Existing query generation methods are limited to generate a fixed number of queries based on this hyper-parameter, which often leads to missed detections and incorrect detections due to the varied number and density of pedestrians in crowded scenes. To address this challenge, we propose an adaptive query generation method called Ranking-based Adaptive Query Generation (RAQG). RAQG comprises three components: a ranking prediction head, a query supplementer, and Soft Gradient L1 Loss (SGL1). Specifically, we leverage the ranking of the lowest confidence score positive training sample to generate queries adaptively. The ranking prediction head predicts this ranking, which guides our query generation. Additionally, to refine the query generation process, we introduce a query supplementer that adjusts the number of queries based on the predicted ranking. Furthermore, we introduce SGL1, a novel loss function for training the ranking prediction head over a wide regression range. Our method is designed to be lightweight and universal, suitable for integration into any DETRs framework for crowded pedestrian detection. Experimental results on Crowdhuman and Citypersons datasets demonstrate that our RAQG method can generate queries adaptively and achieves competitive results. Notably, our approach achieves a state-of-the-art 39.4% MR on Crowdhuman.

CDAN: Convolutional dense attention-guided network for low-light image enhancement

Low-light images, characterized by inadequate illumination, pose challenges of diminished clarity, muted colors, and reduced details. Low-light image enhancement, an essential task in computer vision, aims to rectify these issues by improving brightness, contrast, and overall perceptual quality, thereby facilitating accurate analysis and interpretation. This paper introduces the Convolutional Dense Attention-guided Network (CDAN), a novel solution for enhancing low-light images. CDAN integrates an autoencoder-based architecture with convolutional and dense blocks, complemented by an attention mechanism and skip connections. This architecture ensures efficient information propagation and feature learning. During training, we utilize a composite loss function that merges L2 and VGG losses for improved numeric and perceptual results. Furthermore, a dedicated post-processing phase refines color balance and contrast. Our approach demonstrates notable progress compared to state-of-the-art results in low-light image enhancement, showcasing its robustness across a wide range of challenging scenarios. Our model performs remarkably on benchmark datasets, effectively mitigating under-exposure and proficiently restoring textures and colors in diverse low-light scenarios. This achievement underscores CDAN's potential for diverse computer vision tasks, notably enabling robust object detection and recognition in challenging low-light conditions.

Retinal fundus image super-resolution based on generative adversarial network guided with vascular structure prior

Many ophthalmic and systemic diseases can be screened by analyzing retinal fundus images. The clarity and resolution of retinal fundus images directly determine the effectiveness of clinical diagnosis. Deep learning methods based on generative adversarial networks are used in various research fields due to their powerful generative capabilities, especially image super-resolution. Although Real-ESRGAN is a recently proposed method that excels in processing real-world degraded images, it suffers from structural distortions when super-resolving retinal fundus images are rich in structural information. To address this shortcoming, we first process the input image using a pre-trained U-Net model to obtain a structural segmentation map of the retinal vessels and use the segmentation map as the structural prior. The spatial feature transform layer is then used to better integrate the structural prior into the generation process of the generator. In addition, we introduce channel and spatial attention modules into the skip connections of the discriminator to emphasize meaningful features and accordingly enhance the discriminative power of the discriminator. Based on the original loss functions, we introduce the L1 loss function to measure the pixel-level differences between the segmentation maps of retinal vascular structures in the high-resolution images and the super-resolution images to further constrain the super-resolution images. Simulation results on retinal image datasets show that our improved algorithm results have a better visual performance by suppressing structural distortions in the super-resolution images.

Cross-Modality Image Translation From Brain 18 F-FDG PET/CT Images to Fluid-Attenuated Inversion Recovery Images Using the CypixGAN Framework.

PET/CT and MRI can accurately diagnose dementia but are expensive and inconvenient for patients. Therefore, we aimed to generate synthetic fluid-attenuated inversion recovery (FLAIR) images from 18 F-FDG PET and CT images of the human brain using a generative adversarial network (GAN)-based deep learning framework called the CypixGAN, which combined the CycleGAN framework with the L1 loss function of the pix2pix. Data from 143 patients who underwent PET/CT and MRI were used for training (n = 79), validation (n = 20), and testing (n = 44) the deep learning frameworks. Synthetic FLAIR images were generated using the pix2pix, CycleGAN, and CypixGAN, and white matter hyperintensities (WMHs) were then segmented. The performance of CypixGAN was compared with that of the other frameworks. The CypixGAN outperformed the pix2pix and CycleGAN in generating synthetic FLAIR images with superior visual quality. Peak signal-to-noise ratio and structural similarity index (mean ± standard deviation) estimated using the CypixGAN (20.23 ± 1.31 and 0.80 ± 0.02, respectively) were significantly higher than those estimated using the pix2pix (19.35 ± 1.43 and 0.79 ± 0.02, respectively) and CycleGAN (18.74 ± 1.49 and 0.78 ± 0.02, respectively) ( P < 0.001). WMHs in synthetic FLAIR images generated using the CypixGAN closely resembled those in ground-truth images, as indicated by the low absolute percentage volume differences and high dice similarity coefficients. The CypixGAN generated high-quality FLAIR images owing to the preservation of spatial information despite using unpaired images. This framework may help improve diagnostic performance and cost-effectiveness of PET/CT when MRI scan is unavailable.

Method for Augmenting Side-Scan Sonar Seafloor Sediment Image Dataset Based on BCEL1-CBAM-INGAN.

In this paper, a method for augmenting samples of side-scan sonar seafloor sediment images based on CBAM-BCEL1-INGAN is proposed, aiming to address the difficulties in acquiring and labeling datasets, as well as the insufficient diversity and quantity of data samples. Firstly, a Convolutional Block Attention Module (CBAM) is integrated into the residual blocks of the INGAN generator to enhance the learning of specific attributes and improve the quality of the generated images. Secondly, a BCEL1 loss function (combining binary cross-entropy and L1 loss functions) is introduced into the discriminator, enabling it to focus on both global image consistency and finer distinctions for better generation results. Finally, augmented samples are input into an AlexNet classifier to verify their authenticity. Experimental results demonstrate the excellent performance of the method in generating images of coarse sand, gravel, and bedrock, as evidenced by significant improvements in the Frechet Inception Distance (FID) and Inception Score (IS). The introduction of the CBAM and BCEL1 loss function notably enhances the quality and details of the generated images. Moreover, classification experiments using the AlexNet classifier show an increase in the recognition rate from 90.5% using only INGAN-generated images of bedrock to 97.3% using images augmented using our method, marking a 6.8% improvement. Additionally, the classification accuracy of bedrock-type matrices is improved by 5.2% when images enhanced using the method presented in this paper are added to the training set, which is 2.7% higher than that of the simple method amplification. This validates the effectiveness of our method in the task of generating seafloor sediment images, partially alleviating the scarcity of side-scan sonar seafloor sediment image data.

Image restoration for digital line drawings using line masks

The restoration of digital images holds practical significance due to the fact that degradation of digital image data on the internet is common. State-of-the-art image restoration methods usually employ end-to-end trained networks. However, we argue that a network trained with diverse image pairs is not optimal for restoring line drawings which have extensive plain backgrounds. We propose a line-drawing restoration framework which takes a restoration neural network as backbone and processes an input degraded line drawing in two steps. First, a proposed mask-predicting network predicts a line mask which indicates the possible location of foreground and background in the potential original line drawing. Next, we feed the degraded input line drawing together with the predicted line mask into the backbone restoration network. The traditional L1 loss for the backbone restoration network is substituted with a masked Mean Square Error (MSE) loss. We test our framework on two classical image restoration tasks: JPEG restoration and super-resolution, and experiments demonstrate that our framework can achieve better quantitative and visual results in most cases.

ScRNMF: An imputation method for single-cell RNA-seq data by robust and non-negative matrix factorization.

Single-cell RNA sequencing (scRNA-seq) has emerged as a powerful tool in genomics research, enabling the analysis of gene expression at the individual cell level. However, scRNA-seq data often suffer from a high rate of dropouts, where certain genes fail to be detected in specific cells due to technical limitations. This missing data can introduce biases and hinder downstream analysis. To overcome this challenge, the development of effective imputation methods has become crucial in the field of scRNA-seq data analysis. Here, we propose an imputation method based on robust and non-negative matrix factorization (scRNMF). Instead of other matrix factorization algorithms, scRNMF integrates two loss functions: L2 loss and C-loss. The L2 loss function is highly sensitive to outliers, which can introduce substantial errors. We utilize the C-loss function when dealing with zero values in the raw data. The primary advantage of the C-loss function is that it imposes a smaller punishment for larger errors, which results in more robust factorization when handling outliers. Various datasets of different sizes and zero rates are used to evaluate the performance of scRNMF against other state-of-the-art methods. Our method demonstrates its power and stability as a tool for imputation of scRNA-seq data.

Uncertainty estimation in HDR imaging with Bayesian neural networks

The goal of high dynamic range (HDR) imaging is to estimate potential high-quality images from multi-exposed low dynamic range (LDR) inputs. Intuitively, there exist various possible HDR images corresponding to given LDR inputs, which results in uncertainty in the estimated results. However, most existing HDR imaging methods employ l1 or l2 loss only to provide one possible estimation from various possible solutions, which fails to model the uncertainty, and thus lacks high-frequency details. In this work, we design Bayesian neural networks to capture the uncertainty, which can model one-to-many relations and provide various possible solutions. Concretely, we propose a Variational Bayesian Layer by leveraging a hierarchical prior on the network weights and inferring a new joint posterior, which is utilized to model uncertainty in high-frequency details (e.g., textures), and model uncertainty in semantic information (e.g., ghost areas), respectively. By leveraging Bayesian framework, the proposed method can provide various potential high-quality estimations, especially in high-frequency details. Experiments on different datasets show that the proposed method enables sampling possible HDR imaging, and the consensus estimate achieves state-of-the-art quantitative and qualitative results.

Robust deep learning estimation of cortical bone porosity from MR T1-weighted images for individualized transcranial focused ultrasound planning.

Transcranial focused ultrasound (tFUS) is an emerging neuromodulation approach that has been demonstrated in animals but is difficult to translate to humans because of acoustic attenuation and scattering in the skull. Optimal dose delivery requires subject-specific skull porosity estimates which has traditionally been done using CT. We propose a deep learning (DL) estimation of skull porosity from T1-weighted MRI images which removes the need for radiation-inducing CT scans. We evaluate the impact of different DL approaches, including network architecture, input size and dimensionality, multichannel inputs, data augmentation, and loss functions. We also propose back-propagation in the mask (BIM), a method whereby only voxels inside the skull mask contribute to training. We evaluate the robustness of the best model to input image noise and MRI acquisition parameters and propagate porosity estimation errors in thousands of beam propagation scenarios. Our best performing model is a cGAN with a ResNet-9 generator with 3D 64×64×64 inputs trained with L1 and L2 losses. The model achieved a mean absolute error of 6.9% in the test set, compared to 9.5% with the pseudo-CT of Izquierdo et al. (38% improvement) and 9.4% with the generic pixel-to-pixel image translation cGAN pix2pix (36% improvement). Acoustic dose distributions in the thalamus were more accurate with our approach than with the pseudo-CT approach of both Burgos et al. and Izquierdo et al, resulting in near-optimal treatment planning and dose estimation at all frequencies compared to CT (reference). Our DL approach porosity estimates with ~7% error, is robust to input image noise and MRI acquisition parameters (sequence, coils, field strength) and yields near-optimal treatment planning and dose estimates for both central (thalamus) and lateral brain targets (amygdala) in the 200-1000 kHz frequency range.

DST-DETR: Image Dehazing RT-DETR for Safety Helmet Detection in Foggy Weather.

In foggy weather, outdoor safety helmet detection often suffers from low visibility and unclear objects, hindering optimal detector performance. Moreover, safety helmets typically appear as small objects at construction sites, prone to occlusion and difficult to distinguish from complex backgrounds, further exacerbating the detection challenge. Therefore, the real-time and precise detection of safety helmet usage among construction personnel, particularly in adverse weather conditions such as foggy weather, poses a significant challenge. To address this issue, this paper proposes the DST-DETR, a framework for foggy weather safety helmet detection. The DST-DETR framework comprises a dehazing module, PAOD-Net, and an object detection module, ST-DETR, for joint dehazing and detection. Initially, foggy images are restored within PAOD-Net, enhancing the AOD-Net model by introducing a novel convolutional module, PfConv, guided by the parameter-free average attention module (PfAAM). This module enables more focused attention on crucial features in lightweight models, therefore enhancing performance. Subsequently, the MS-SSIM + ℓ2 loss function is employed to bolster the model's robustness, making it adaptable to scenes with intricate backgrounds and variable fog densities. Next, within the object detection module, the ST-DETR model is designed to address small objects. By refining the RT-DETR model, its capability to detect small objects in low-quality images is enhanced. The core of this approach lies in utilizing the variant ResNet-18 as the backbone to make the network lightweight without sacrificing accuracy, followed by effectively integrating the small-object layer into the improved BiFPN neck structure, resulting in CCFF-BiFPN-P2. Various experiments were conducted to qualitatively and quantitatively compare our method with several state-of-the-art approaches, demonstrating its superiority. The results validate that the DST-DETR algorithm is better suited for foggy safety helmet detection tasks in construction scenarios.

Robust pixel-wise illuminant estimation algorithm for images with a low bit-depth.

Conventional illuminant estimation methods were developed for scenes with a uniform illumination, while recently developed methods, such as pixel-wise methods, estimate the illuminants at the pixel level, making them applicable to a wider range of scenes. It was found that the same pixel-wise algorithm had very different performance when applied to images with different bit-depths, with up to a 30% decrease in accuracy for images having a lower bit-depth. Image signal processing (ISP) pipelines, however, prefer to deal with images with a lower bit-depth. In this paper, the analyses show that such a reduction was due to the loss of details and increase of noises, which were never identified in the past. We propose a method combining the L1 loss optimization and physical-constrained post-processing. The proposed method was found to result in around 40% higher estimation accuracy, in comparison to the state-of-the-art DNN-based methods.

Influence of loss function and electron dose on ptychography of 2D materials using the Wirtinger flow

Iterative phase retrieval is based on minimising a loss function as a measure of the consistency of an initial guess and underlying experimental data. Under ideal experimental conditions, real data contains Poissonian noise due to counting statistics. In this work, we use the Wirtinger Flow concept in combination with four common loss functions, being the L1 loss, the mean-squared error (MSE), the amplitude loss and the Poisson loss. Since only the latter reflects the counting statistics as an asymmetric Poisson distribution correctly, our simulation study focuses on two main cases. Firstly, high-dose momentum-resolved scanning transmission electron microscopy (STEM) of an MoS2 monolayer is considered for phase retrieval. In this case, it is found that the four losses perform differently with respect to chemical sensitivity and frequency transfer, which we interprete in terms of the substantially different signal level in the bright and dark field part of diffraction patterns. Remedies are discussed using further simulations, addressing the use of virtual ring detectors for the dark field, or restricting loss calculation to the bright field. Secondly, a dose series is presented down to 100 electrons per diffraction pattern. It is found that all losses yield qualitatively reasonable structural data in the phase, whereas only MSE and Poisson loss range at the correct amplitude level. Chemical contrast is, in general, reliably obtained using the Poisson concept, which also provides the most continuous spatial frequency transfer as to the reconstructed object transmission function.

Integrated Neural Network Approach for Enhanced Vital Signal Analysis Using CW Radar

This study introduces a novel approach for analyzing vital signals using continuous-wave (CW) radar, employing an integrated neural network model to overcome the limitations associated with traditional step-by-step signal processing methods. Conventional methods for vital signal monitoring, such as electrocardiograms (ECGs) and sphygmomanometers, require direct contact and impose constraints on specific scenarios. Conversely, our study primarily focused on non-contact measurement techniques, particularly those using CW radar, which is known for its simplicity but faces challenges such as noise interference and complex signal processing. To address these issues, we propose a temporal convolutional network (TCN)-based framework that seamlessly integrates noise removal, demodulation, and fast Fourier transform (FFT) processes into a single neural network. This integration minimizes cumulative errors and processing time, which are common drawbacks of conventional methods. The TCN was trained using a dataset comprising preprocessed in-phase and quadrature (I/Q) signals from the CW radar and corresponding heart rates measured via ECG. The performance of the proposed method was evaluated based on the L1 loss and accuracy against the moving average of the estimated heart rates. The results indicate that the proposed approach has the potential for efficient and accurate non-contact vital signal analysis, opening new avenues in health monitoring and medical research. Additionally, the integration of CW radar and neural networks in our framework offers a robust and scalable solution, enhancing the practicality of non-contact health monitoring systems in diverse environments. This technology can be leveraged in healthcare robots to provide continuous and unobtrusive monitoring of patients’ vital signs, enabling timely interventions and improving overall patient care.

A multiscale 3D network for lung nodule detection using flexible nodule modeling.

Lung cancer is the most common type of cancer. Detection of lung cancer at an early stage can reduce mortality rates. Pulmonary nodules may represent early cancer and can be identified through computed tomography (CT) scans. Malignant risk can be estimated based on attributes like size, shape, location, and density. Deep learning algorithms have achieved remarkable advancements in this domain compared to traditional machine learning methods. Nevertheless, many existing anchor-based deep learning algorithms exhibit sensitivity to predefined anchor-box configurations, necessitating manual adjustments to obtain optimal outcomes. Conversely, current anchor-free deep learning-based nodule detection methods normally adopt fixed-size nodule models like cubes or spheres. To address these technical challenges, we propose a multiscale 3D anchor-free deep learning network (M3N) for pulmonary nodule detection, leveraging adjustable nodule modeling (ANM). Within this framework, ANM empowers the representation of target objects in an anisotropic manner, with a novel point selection strategy (PSS) devised to accelerate the learning process of anisotropic representation. We further incorporate a composite loss function that combines the conventional L2 loss and cosine similarity loss, facilitating M3N to learn nodules' intensity distribution in three dimensions. Experiment results show that the M3N achieves 90.6% competitive performance metrics (CPM) with seven predefined false positives per scan on the LUNA 16 dataset. This performance appears to exceed that of other state-of-the-art deep learning-based networks reported in their respective publications. Individual test results also demonstrate that M3N excels in providing more accurate, adaptive bounding boxes surrounding the contours of target nodules. The newly developed nodule detection system reduces reliance on prior knowledge, such as the general size of objects in the dataset, thus it should enhance overall robustness and versatility. Distinct from traditional nodule modeling techniques, the ANM approach aligns more closely with the morphological characteristics of nodules. Time consumption and detection results demonstrate promising efficiency and accuracy which should be validated in clinical settings.

Data augmentation for Gram-stain images based on Vector Quantized Variational AutoEncoder

Availability of large-scale datasets plays a significant role in segmentation and classification tasks using deep learning. However, domains such as healthcare inherently suffer from unavailability and inaccessibility of data. This leads to challenges in deploying CNN-based models for computer-aided diagnosis. This challenge extends to Gram-stain image analysis for detecting bacterial infections, which is a crucial task. The lack of datasets containing Gram-stained direct and culture smear images exacerbates the significant challenges in deep learning tasks. In this regard, we investigate a novel application of the Variational AutoEncoder. Specifically, the Vector Quantized Variational AutoEncoder model is trained to generate the Gram-stain images. Incorporating a novel loss function, where the quality loss (Lqu) is derived by integrating the LossSSIM, L1, and L2 losses with the VQ-VAE loss (Lossvq) for proposed approach for Gram-stained direct and culture smear images. This modification facilitates the creation of images closely resembling the original input, leading to notable SSIM scores of 0.92 for Gram-stained culture images and 0.88 for Gram-stained direct smear images. The current study compares the proposed method with state-of-the-art machine learning based and CNN based transformations. This work also demonstrates the classification process with and without image augmentation. It shows that the area under the curve in the case of augmentation is higher by an average of 20%.

An improved infrared simulation method based on generative adversarial networks

Infrared simulation technology has great application value in the fields of UAV target simulation and infrared signal processing. However, traditional infrared simulation technology has the disadvantages of poor scene adaptability, low fidelity of infrared simulation images and long training time. This paper proposes an infrared image simulation algorithm based on improved Generative Adversarial Networks(GAN) to generate multi-scene infrared images, which could provide a reliable expansion of infrared database. First, we construct a GAN model by introducing a multi-resolution hierarchical structure and attention mechanism, which can effectively enhance the model’s adaptability to different scenes during the simulation of generating infrared images. Secondly, the SSIM loss term and L1 loss term are introduced during the training process to form a composite loss function, which prompts the proposed GAN to generate high-fidelity infrared simulation images. Finally, while maintaining the quality of simulated infrared images, a lightweight model is designed through model compression to significantly reduce the number of model parameters.The traditional discriminator based on the UNet-self attention is replaced by a discriminator based on attention mechanisms and adaptive max-pooling structure. The proposed GAN is trained and tested on three public datasets. Experimental results demonstrate that the proposed method improves structural similarity by 11% compared to Pix2Pix and achieves a Peak Signal-to-Noise Ratio (PSNR) of 25.8, which is 3.9 higher than InfraGAN. At the same time, compared with InfraGAN, the number of model parameters is reduced by 68% , which makes the method efficiently obtain simulated infrared images corresponding to given visible images.