Recent research on learning a mapping between raw Bayer images and RGB images has progressed with the development of deep convolutional neural network. A challenging data set namely the Zurich Raw-to-RGB data set (ZRR) has been released in the AIM 2019 raw-to-RGB mapping challenge. In ZRR, input raw and target RGB images are captured by two different cameras and thus not perfectly aligned. Moreover, camera metadata such as white balance gains and color correction matrix are not provided, which makes the challenge more difficult. In this paper, we explore an effective network structure and a loss function to address these issues. We exploit a two-stage U-Net architecture, and also introduce a loss function that is less variant to alignment and more sensitive to color differences. In addition, we show an ensemble of networks trained with different loss functions can bring a significant performance gain. We demonstrate the superiority of our method by achieving the highest score in terms of both the peak signal-to-noise ratio and the structural similarity and obtaining the second-best mean-opinion-score in the challenge.

X-ray baggage inspection has been widely used for maintaining airport and transportation security. Towards automated inspection, recent deep learning-based methods have attempted to detect hazardous objects directly from X-ray images. Since it is challenging to collect a large number of training images from real-world environments, most previous learning-based methods rely on image synthesis for training data generation. However, these methods randomly combine foreground and background images, restricting the effectiveness of synthetic images for object detection. To solve this problem, in this paper, we propose a learning-based X-ray image synthesis method for object detection. Specifically, for each foreground object to be synthesized, we first estimate positions difficult to detect by the object detector. These positions and their corresponding confidence values are then used to construct a difficulty map, which is used for sampling the target foreground position for image synthesis. The performance analysis using various state-of-the-art object detectors shows that the proposed synthesis method can produce more useful training data compared with the conventional random synthesis method.INDEX TERMS Deep learning, neural network, object detection, X-ray, inspection.

Lossy video compression achieves coding gains at the expense of the quality loss of the decoded images. Owing to the success of deep learning techniques, especially convolutional neural networks (CNNs), many compression artifacts reduction (CAR) techniques have been used to significantly improve the quality of decoded images by applying CNNs which are trained to predict the original artifactfree images from the decoded images. Most existing video compression standards control the compression ratio using a quantization parameter (QP), so the quality of the decoded images is strongly QP-dependent. Training individual CNNs for predetermined QPs is one of the common approaches to dealing with different levels of compression artifacts. However, compression artifacts are also dependent on the local characteristics of an image. Therefore, a CNN trained for specific QP cannot fully remove the compression artifacts of all images, even those encoded using the same QP. In this paper, we introduce a pixel-precise network selection network (PNSNet). From multiple reconstructed images obtained using multiple QPspecific CAR networks, PNSNet is trained to find the best CAR network for each pixel. The output of PNSNet is then used as an explicit spatial attention channel for an image fusion network that combines multiple reconstructed images. Experimental results demonstrated that the quality of decoded images can be significantly improved by the proposed multiple CAR network fusion method.INDEX TERMS Compression artifacts reduction, convolutional neural network, deep learning, network fusion, video compression.

No abstract

Images captured from real-world environments often include blur artifacts resulting from camera movement, dynamic object motion, or out-of-focus. Although such blur artifacts are inevitable, most object detection methods do not have special considerations for them; therefore, they may fail to detect objects in blurry images. One possible solution is applying image deblurring prior to object detection. However, this solution is computationally demanding and its performance heavily depends on image deblurring results. In this study, we propose a novel blur-aware object detection framework. First, we construct a synthetic but realistic dataset by applying a diverse set of motion blur kernels to blur-free images. Subsequently, we leverage self-guided knowledge distillation between the teacher and student networks that perform object detection using blur-free and blurry images, respectively. The teacher and student networks share most of their network parameters and jointly learn in a fully-supervised manner. The teacher network provides image features as hints for feature-level deblurring and also renders soft labels for the training of the student network. Guided by the hints and the soft labels from the teacher, the student network learns and expands their knowledge on object detection in blurry images. Experimental results show that the proposed framework improves the robustness of several widely used object detectors against image blurs.