13 July 2020 in IEEE Transactions on Image Processing
Single image defogging is a classical and challenging problem in computer vision. Existing methods towards this problem mainly include handcrafted priors based methods that rely on the use of the atmospheric degradation model and learning-based approaches that require paired fog-fogfree training example images. In practice, however, prior-based methods are prone to failure due to their own limitations and paired training data are extremely difficult to acquire. Moreover, there are few studies on the unpaired trainable defogging network in this field. Thus, inspired by the principle of CycleGAN network, we have developed an end-to-end learning system that uses unpaired fog and fogfree training images, adversarial discriminators and cycle consistency losses to automatically construct a fog removal system. Similar to CycleGAN, our system has two transformation paths; one maps fog images to a fogfree image domain and the other maps fogfree images to a fog image domain. Instead of one stage mapping, our system uses a two stage mapping strategy in each transformation path to enhance the effectiveness of fog removal. Furthermore, we make explicit use of prior knowledge in the networks by embedding the atmospheric degradation principle and a sky prior for mapping fogfree images to the fog images domain. In addition, we also contribute the first real world nature fog-fogfree image dataset for defogging research. Our multiple real fog images dataset (MRFID) contains images of 200 natural outdoor scenes. For each scene, there is one clear image and corresponding four foggy images of different fog densities manually selected from a sequence of images taken by a fixed camera over the course of one year. Qualitative and quantitative comparison against several state-of-the-art methods on both synthetic and real world images demonstrate that our approach is effective and performs favorably for recovering a clear image from a foggy image.
15 November 2018 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
This paper presents an advanced urban traffic density estimation solution using the latest deep learning techniques to intelligently process ultrahigh-resolution traffic videos taken from an unmanned aerial vehicle (UAV). We first capture nearly an hour-long ultrahigh-resolution traffic video at five busy road intersections of a modern megacity by flying a UAV during the rush hours. We then randomly sampled over 17 K 512 × 512 pixel image patches from the video frames and manually annotated over 64 K vehicles to form a dataset for this paper, which will also be made available to the research community for research purposes. Our innovative urban traffics analysis solution consists of an advanced deep neural network (DNN) based vehicle detection and localization, type (car, bus, and truck) recognition, tracking, and vehicle counting over time. We will present extensive experimental results to demonstrate the effectiveness of our solution. We will show that our enhanced single shot multibox detector (Enhanced-SSD) outperforms other DNN-based techniques and that deep learning techniques are more effective than traditional computer vision techniques in traffic video analysis. We will also show that ultrahigh-resolution video provides more information that enables more accurate vehicle detection and recognition than lower resolution contents. This paper not only demonstrates the advantages of using the latest technological advancements (ultrahigh-resolution video and UAV), but also provides an advanced DNN-based solution for exploiting these technological advancements for urban traffic density estimation.
24 December 2019 in IEEE Transactions on Medical Imaging
Deep learning approaches are widely applied to histopathological image analysis due to the impressive levels of performance achieved. However, when dealing with high-resolution histopathological images, utilizing the original image as input to the deep learning model is computationally expensive, while resizing the original image to achieve low resolution incurs information loss. Some hard-attention based approaches have emerged to select possible lesion regions from images to avoid processing the original image. However, these hard-attention based approaches usually take a long time to converge with weak guidance, and valueless patches may be trained by the classifier. To overcome this problem, we propose a deep selective attention approach that aims to select valuable regions in the original images for classification. In our approach, a decision network is developed to decide where to crop and whether the cropped patch is necessary for classification. These selected patches are then trained by the classification network, which then provides feedback to the decision network to update its selection policy. With such a co-evolution training strategy, we show that our approach can achieve a fast convergence rate and high classification accuracy. Our approach is evaluated on a public breast cancer histopathological image database, where it demonstrates superior performance compared to state-of-the-art deep learning approaches, achieving approximately 98% classification accuracy while only taking 50% of the training time of the previous hard-attention approach.
1 May 2019 in Neurocomputing
17 March 2022 in IEEE Transactions on Multimedia
As a sub-topic of Text-to-Image synthesis, Text-to-Face generation has a great potential in face related applications. In this paper, we propose a generic Text-to-Face framework, namely TextFace, to achieve diverse and high-quality face image generation from text description. We introduce a novel method called Text-to-Style mapping, where the text description can be directly encoded into the latent space of a pretrained StyleGAN. Guided by our text-image similarity matching and face captioning based text alignment, the textual latent code can be fed into a well-trained StyleGAN's generator, to produce diverse face images with high resolution (1024 1024). Furthermore, our model inherently supports the semantic face editing using text descriptions. Finally, experimental results quantitatively and qualitatively demonstrate the superior performance of our model.
23 September 2024 in ACM Transactions on Multimedia Computing, Communications, and Applications
Low-Light Image Enhancement (LLIE) presents challenges due to texture information loss and uneven illumination, which can distort feature distribution and reduce the quality of the enhanced images. However, current deep learning methods for LLIE only use supervised information from clear images to extract low-light image features, while disregarding the negative information in low-light images (i.e., low illumination and noise). To address these challenges, we propose a novel LLIE method, LACR-VAE, by leveraging the negative information and considering the uneven illumination. In particular, a Light-Aware Contrastive Regularization (LACR) based on contrastive learning is designed to exploit information from both clear and low-light images. The LACR aims to align latent variables of enhanced images with clear images, away from those of low-light images. This allows the method to prioritize essential elements for LLIE and minimize noise and lighting variations. Furthermore, considering the uneven illumination with diverse region sizes and shapes, a Region-CAlibrated Module (RCAM) is present to learn local and global illumination relations among image regions, and an Attention-guided Multi-Scale Module (AMSM) is designed to extract multi-scale features that improve the model’s representation capability. Extensive experiments show that our method achieves superior performance than previous works. Specifically, our method yields a significant enhancement in the National Aeronautics and Space Administration (NASA) testset, achieving an improvement of at least 0.99 in PSNR and 0.0409 in SSIM. Codes and datasets are available at https://github.com/csxuwu/LACR-VAE .