Predicting the grasping point accurately and quickly is crucial for successful robotic manipulation. However, to commercially deploy a robot, such as a dishwasher robot in a commercial kitchen, we also need to consider the constraints of limited usable resources. We present a deep learning method to predict the grasp position when using a single suction gripper for picking up objects. The proposed method is based on a shallow network to enable lower training costs and efficient inference on limited resources. Costs are further reduced by collecting data in a custom-built synthetic environment. For evaluating the proposed method, we developed a system that models a commercial kitchen for a dishwasher robot to manipulate symmetric objects. We tested our method against a model-fitting method and an algorithm-based method in our developed commercial kitchen environment and found that a shallow network trained with only the synthetic data achieves high accuracy. We also demonstrate the practicality of using a shallow network in sequence with an object detector for ease of training, prediction speed, low computation cost, and easier debugging.

In the neck part of a two-stage object detection network, feature fusion is generally carried out in either a top-down or bottom-up manner. However, two types of imbalance may exist: feature imbalance in the neck of the model and gradient imbalance in the region of interest extraction layer due to the scale changes of objects. The deeper the network is, the more abstract the learned features are, that is to say, more semantic information can be extracted. However, the extracted image background, spatial location, and other resolution information are less. In contrast, the shallow part can learn little semantic information, but a lot of spatial location information. We propose the Both Ends to Centre to Multiple Layers (BEtM) feature fusion method to solve the feature imbalance problem in the neck and a Multi-level Region of Interest Feature Extraction (MRoIE) layer to solve the gradient imbalance problem. In combination with the Region-based Convolutional Neural Network (R-CNN) framework, our Balanced Feature Fusion (BFF) method offers significantly improved network performance compared with the Faster R-CNN architecture. On the MS COCO 2017 dataset, it achieves an average precision (AP) that is 1.9 points and 3.2 points higher than those of the Feature Pyramid Network (FPN) Faster R-CNN framework and the Generic Region of Interest Extractor (GRoIE) framework, respectively.

Loan default prediction has been a significant problem in the financial domain because overdue loans may incur significant losses. Machine learning methods have been introduced to solve this problem, but there are still many challenges including feature multicollinearity, imbalanced labels, and small data sample problems. To replicate the success of deep learning in many areas, an effective regularization technique named muddling label regularization is introduced in this letter, and an ensemble of feed-forward neural networks is proposed, which outperforms machine learning and deep learning baselines in a real-world dataset.

As the basis of fine breeding management and animal husbandry insurance, individual recognition of dairy cattle is an important issue in the animal husbandry management field. Due to the limitations of the traditional method of cow identification, such as being easy to drop and falsify, it can no longer meet the needs of modern intelligent pasture management. In recent years, with the rise of computer vision technology, deep learning has developed rapidly in the field of face recognition. The recognition accuracy has surpassed the level of human face recognition and has been widely used in the production environment. However, research on the facial recognition of large livestock, such as dairy cattle, needs to be developed and improved. According to the idea of a residual network, an improved convolutional neural network (Res_5_2Net) method for individual dairy cow recognition is proposed based on dairy cow facial images in this letter. The recognition accuracy on our self-built cow face database (3012 training sets, 1536 test sets) can reach 94.53%. The experimental results show that the efficiency of identification of dairy cows is effectively improved.

To accelerate deep learning (DL) processes on the supercomputer Fugaku, the authors have ported and optimized oneDNN for Fugaku's CPU, the Fujitsu A64FX. oneDNN is an open-source DL processing library developed by Intel for the x86_64 architecture. The A64FX CPU is based on the Armv8-A architecture. oneDNN dynamically creates the execution code for the computation kernels, which are implemented at the granularity of x86_64 instructions using Xbyak, the Just-In-Time (JIT) assembler for x86_64 architecture. To port oneDNN to A64FX, it must be rewritten into Armv8-A instructions using Xbyak_aarch64, the JIT assembler for the Armv8-A architecture. This is challenging because the number of steps to be rewritten exceeds several tens of thousands of lines. This study presents the Xbyak_translator_aarch64. Xbyak_translator_aarch64 is a binary translator that at runtime converts dynamically produced executable codes for the x86_64 architecture into executable codes for the Armv8-A architecture. Xbyak_translator_aarch64 eliminates the need to rewrite the source code for porting oneDNN to A64FX and allows us to port oneDNN to A64FX quickly.

Most existing object detection methods and text detection methods are mainly designed to detect either text or objects. In some scenarios where the task is to find the target word pointed-at by an object, results of existing methods are far from satisfying. However, such scenarios happen often in human-computer interaction, when the computer needs to figure out which word the user is pointing at. Comparing with object detection, pointed-at word localization (PAWL) requires higher accuracy, especially in dense text scenarios. Moreover, in printed document, characters are much smaller than those in scene text detection datasets such as ICDAR-2013, ICDAR-2015 and ICPR-2018 etc. To address these problems, the authors propose a novel target word localization network (TWLN) to detect the pointed-at word in printed documents. In this work, a single deep neural network is trained to extract the features of markers and text sequentially. For each image, the location of the marker is predicted firstly, according to the predicted location, a smaller image is cropped from the original image and put into the same network, then the location of pointed-at word is predicted. To train and test the networks, an efficient approach is proposed to generate the dataset from PDF format documents by inserting markers pointing at the words in the documents, which avoids laborious labeling work. Experiments on the proposed dataset demonstrate that TWLN outperforms the compared object detection method and optical character recognition method on every category of targets, especially when the target is a single character that only occupies several pixels in the image. TWLN is also tested with real photographs, and the accuracy shows no significant differences, which proves the validity of the generating method to construct the dataset.

Previous studies on anomaly detection in videos have trained detectors in which reconstruction and prediction tasks are performed on normal data so that frames on which their task performance is low will be detected as anomalies during testing. This paper proposes a new approach that involves sorting video clips, by using a generative network structure. Our approach learns spatial contexts from appearances and temporal contexts from the order relationship of the frames. Experiments were conducted on four datasets, and we categorized the anomalous sequences by appearance and motion. Evaluations were conducted not only on each total dataset but also on each of the categories. Our method improved detection performance on both anomalies with different appearance and different motion from normality. Moreover, combining our approach with a prediction method produced improvements in precision at a high recall.

Direction of arrival (DOA) estimation of wireless signals is demanded in many applications. In addition to classical methods such as MUSIC and ESPRIT, non-linear algorithms such as compressed sensing have become common subjects of study recently. Deep learning or machine learning is also known as a non-linear algorithm and has been applied in various fields. Generally, DOA estimation using deep learning is classified as on-grid estimation. A major problem of on-grid estimation is that the accuracy may be degraded when the DOA is near the boundary. To reduce such estimation errors, we propose a method of combining two DNNs whose grids are offset by one half of the grid size. Simulation results show that our proposal outperforms MUSIC which is a typical off-grid estimation method. Furthermore, it is shown that the DNN specially trained for a close DOA case achieves very high accuracy for that case compared with MUSIC.

High Frequency Surface Wave Radar (HFSWR) can achieve over-the-horizon detection, which can effectively detect and track the ships and ultra-low altitude aircrafts, as well as the acquisition of sea state information such as icebergs and ocean currents and so on. However, HFSWR is seriously affected by the clutters, especially sea clutter and ionospheric clutter. In this paper, we propose a deep learning image semantic segmentation method based on optimized Deeplabv3+ network to achieve the automatic detection of sea clutter and ionospheric clutter using the measured R-D spectrum images of HFSWR during the typhoon as experimental data, which avoids the disadvantage of traditional detection methods that require a large amount of a priori knowledge and provides a basis for subsequent the clutter suppression or the clutter characteristics research.

There have been some researchers that investigate the accuracy of the approximation to a function that shows a generating pattern of data by a deep neural network. However, they have confirmed only whether at least one function close to the function showing a generating pattern exists in function classes of deep neural networks whose parameter values are changing. Therefore, we propose a new criterion to infer the approximation accuracy. Our new criterion shows the existence ratio of functions close to the function showing a generating pattern in the function classes. Moreover, we show a deep neural network with a larger number of layers approximates the function showing a generating pattern more accurately than one with a smaller number of layers under the proposed criterion, with numerical simulations.

In the recent years, deep learning has achieved significant results in various areas of machine learning. Deep learning requires a huge amount of data to train a model, and data collection techniques such as web crawling have been developed. However, there is a risk that these data collection techniques may generate incorrect labels. If a deep learning model for image classification is trained on a dataset with noisy labels, the generalization performance significantly decreases. This problem is called Learning with Noisy Labels (LNL). One of the recent researches on LNL, called DivideMix [1], has successfully divided the dataset into samples with clean labels and ones with noisy labels by modeling loss distribution of all training samples with a two-component Mixture Gaussian model (GMM). Then it treats the divided dataset as labeled and unlabeled samples and trains the classification model in a semi-supervised manner. Since the selected samples have lower loss values and are easy to classify, training models are in a risk of overfitting to the simple pattern during training. To train the classification model without overfitting to the simple patterns, we propose to introduce consistency regularization on the selected samples by GMM. The consistency regularization perturbs input images and encourages model to outputs the same value to the perturbed images and the original images. The classification model simultaneously receives the samples selected as clean and their perturbed ones, and it achieves higher generalization performance with less overfitting to the selected samples. We evaluated our method with synthetically generated noisy labels on CIFAR-10 and CIFAR-100 and obtained results that are comparable or better than the state-of-the-art method.

Neural network-based three-dimensional (3D) reconstruction methods have produced promising results. However, they do not pay particular attention to reconstructing detailed parts of objects. This occurs because the network is not designed to capture the fine details of objects. In this paper, we propose a network designed to capture both the coarse and fine details of objects to improve the reconstruction of the fine parts of objects.

Since deep learning workloads perform a large number of matrix operations on training data, GPUs (Graphics Processing Units) are efficient especially for the training phase. A cluster of computers each of which equips multiple GPUs can significantly accelerate the deep learning workloads. More specifically, a back-propagation algorithm following a gradient descent approach is used for the training. Although the gradient computation is still a major bottleneck of the training, gradient aggregation and optimization impose both communication and computation overheads, which should also be reduced for further shortening the training time. To address this issue, in this paper, multiple GPUs are interconnected with a PCI Express (PCIe) over 10Gbit Ethernet (10GbE) technology. Since these remote GPUs are interconnected with network switches, gradient aggregation and optimizers (e.g., SGD, AdaGrad, Adam, and SMORMS3) are offloaded to FPGA-based 10GbE switches between remote GPUs; thus, the gradient aggregation and parameter optimization are completed in the network. The proposed FPGA-based 10GbE switches with the four optimizers are implemented on NetFPGA-SUME board. Their resource utilizations are increased by PEs for the optimizers, and they consume up to 56% of the resources. Evaluation results using four remote GPUs connected via the proposed FPGA-based switch demonstrate that these optimizers are accelerated by up to 3.0x and 1.25x compared to CPU and GPU implementations, respectively. Also, the gradient aggregation throughput by the FPGA-based switch achieves up to 98.3% of the 10GbE line rate.

In this paper, we present a simple yet powerful deep neural network for natural image dehazing. The proposed method is designed based on U-Net architecture and we made some design changes to make it better. We first use Group Normalization to replace Batch Normalization to solve the problem of insufficient batch size due to hardware limitations. Second, we introduce FReLU activation into the U-Net block, which can achieve capturing complicated visual layouts with regular convolutions. Experimental results on public benchmarks demonstrate the effectiveness of the modified components. On the SOTS Indoor and Outdoor datasets, it obtains PSNR of 32.23 and 31.64 respectively, which are comparable performances with state-of-the-art methods. The code is publicly available online soon.

We propose deep neural network (DNN)-based speech enhancement that reduces musical noise and achieves better auditory impressions. The musical noise is an artifact generated by nonlinear signal processing and negatively affects the auditory impressions. We aim to develop musical-noise-free speech enhancement methods that suppress the musical noise generation and produce perceptually-comfortable enhanced speech. DNN-based speech enhancement using a soft mask achieves high noise reduction but generates musical noise in non-speech regions. Therefore, first, we define kurtosis matching for DNN-based low-musical-noise speech enhancement. Kurtosis is the fourth-order moment and is known to correlate with the amount of musical noise. The kurtosis matching is a penalty term of the DNN training and works to reduce the amount of musical noise. We further extend this scheme to standardized-moment matching. The extended scheme involves using moments whose orders are higher than kurtosis and generalizes the conventional musical-noise-free method based on kurtosis matching. We formulate standardized-moment matching and explore how effectively the higher-order moments reduce the amount of musical noise. Experimental evaluation results 1) demonstrate that kurtosis matching can reduce musical noise without negatively affecting noise suppression and 2) newly reveal that the sixth-moment matching also achieves low-musical-noise speech enhancement as well as kurtosis matching.

We propose a new framework for estimating depth information from a single image. Our framework is relatively small and straightforward by employing a two-stage architecture: a residual network and a simple decoder network. Our residual network in this paper is a remodeled of the original ResNet-50 architecture, which consists of only thirty-eight convolution layers in the residual block following by pair of two up-sampling and layers. While the simple decoder network, stack of five convolution layers, accepts the initial depth to be refined as the final output depth. During training, we monitor the loss behavior and adjust the learning rate hyperparameter in order to improve the performance. Furthermore, instead of using a single common pixel-wise loss, we also compute loss based on gradient-direction, and their structure similarity. This setting in our network can significantly reduce the number of network parameters, and simultaneously get a more accurate image depth map. The performance of our approach has been evaluated by conducting both quantitative and qualitative comparisons with several prior related methods on the publicly NYU and KITTI datasets.

Rapid building damage identification plays a vital role in rescue operations when disasters strike, especially when rescue resources are limited. In the past years, supervised machine learning has made considerable progress in building damage identification. However, the usage of supervised machine learning remains challenging due to the following facts: 1) the massive samples from the current damage imagery are difficult to be labeled and thus cannot satisfy the training requirement of deep learning, and 2) the similarity between partially damaged and undamaged buildings is high, hindering accurate classification. Leveraging the abundant samples of auxiliary domains, domain adaptation aims to transfer a classifier trained by historical damage imagery to the current task. However, traditional domain adaptation approaches do not fully consider the category-specific information during feature adaptation, which might cause negative transfer. To address this issue, we propose a novel domain adaptation framework that individually aligns each category of the target domain to that of the source domain. Our method combines the variational autoencoder (VAE) and the Gaussian mixture model (GMM). First, the GMM is established to characterize the distribution of the source domain. Then, the VAE is constructed to extract the feature of the target domain. Finally, the Kullback-Leibler (KL) divergence is minimized to force the feature of the target domain to observe the GMM of the source domain. Two damage detection tasks using post-earthquake and post-hurricane imageries are utilized to verify the effectiveness of our method. Experiments show that the proposed method obtains improvements of 4.4% and 9.5%, respectively, compared with the conventional method.

Video-based person re-identification (re-ID) aims at retrieving person across non-overlapping camera and has achieved promising results owing to deep convolutional neural network. Due to the dynamic properties of the video, the problems of background clutters and occlusion are more serious than image-based person Re-ID. In this letter, we present a novel triple attention network (TriANet) that simultaneously utilizes temporal, spatial, and channel context information by employing the self-attention mechanism to get robust and discriminative feature. Specifically, the network has two parts, where the first part introduces a residual attention subnetwork, which contains channel attention module to capture cross-dimension dependencies by using rotation and transformation and spatial attention module to focus on pedestrian feature. In the second part, a time attention module is designed to judge the quality score of each pedestrian, and to reduce the weight of the incomplete pedestrian image to alleviate the occlusion problem. We evaluate our proposed architecture on three datasets, iLIDS-VID, PRID2011 and MARS. Extensive comparative experimental results show that our proposed method achieves state-of-the-art results.

In this study, we propose a method to estimate body composition-related health indicators (e.g., ratio of body fat, body water, and muscle, etc.) using video-based gait analysis. This method is more efficient than individual measurement using a conventional body composition meter. Specifically, we designed a deep-learning framework with a convolutional neural network (CNN), where the input is a gait energy image (GEI) and the output consists of the health indicators. Although a vast amount of training data is typically required to train network parameters, it is unfeasible to collect sufficient ground-truth data, i.e., pairs consisting of the gait video and the health indicators measured using a body composition meter for each subject. We therefore use a two-step approach to exploit an auxiliary gait dataset that contains a large number of subjects but lacks the ground-truth health indicators. At the first step, we pre-train a backbone network using the auxiliary dataset to output gait primitives such as arm swing, stride, the degree of stoop, and the body width — considered to be relevant to the health indicators. At the second step, we add some layers to the backbone network and fine-tune the entire network to output the health indicators even with a limited number of ground-truth data points of the health indicators. Experimental results show that the proposed method outperforms the other methods when training from scratch as well as when using an auto-encoder-based pre-training and fine-tuning approach; it achieves relatively high estimation accuracy for the body composition-related health indicators except for body fat-relevant ones.

Binary neural networks (BNNs), where both activations and weights are radically quantized to be {-1, +1}, can massively accelerate the run-time performance of convolution neural networks (CNNs) for edge devices, by computation complexity reduction and memory footprint saving. However, the non-differentiable binarizing function used in BNNs, makes the binarized models hard to be optimized, and introduces significant performance degradation than the full-precision models. Many previous works managed to correct the backward gradient of binarizing function with various improved versions of straight-through estimation (STE), or in a gradual approximate approach, but the gradient suppression problem was not analyzed and handled. Thus, we propose a novel gradient corrected approximation (GCA) method to match the discrepancy between binarizing function and backward gradient in a gradual and stable way. Our work has two primary contributions: The first is to approximate the backward gradient of binarizing function using a simple leaky-steep function with variable window size. The second is to correct the gradient approximation by standardizing the backward gradient propagated through binarizing function. Experiment results show that the proposed method outperforms the baseline by 1.5% Top-1 accuracy on ImageNet dataset without introducing extra computation cost.