Search

Simple Summary: Identifying bird species is crucial in various bird monitoring tasks. In this study, we implemented a bird species recognition method based on visual features by convolutional neural networks. This paper proposed an attention-guided data enhancement method and a decoupled knowledge distillation model compression method. Using the two methods, we created a novel and efficient lightweight fine-grained bird classification model. The model not only achieved high accuracy in bird classification but was also user-friendly for edge devices such as mobile. Our work may be useful for bird species recognition and bird biodiversity monitoring. Classifying birds accurately is essential for ecological monitoring. In recent years, bird image classification has become an emerging method for bird recognition. However, the bird image classification task needs to face the challenges of high intraclass variance and low inter-class variance among birds, as well as low model efficiency. In this paper, we propose a fine-grained bird classification method based on attention and decoupled knowledge distillation. First of all, we propose an attention-guided data augmentation method. Specifically, the method obtains images of the object's key part regions through attention. It enables the model to learn and distinguish fine features. At the same time, based on the localization–recognition method, the bird category is predicted using the object image with finer features, which reduces the influence of background noise. In addition, we propose a model compression method of decoupled knowledge distillation. We distill the target and nontarget class knowledge separately to eliminate the influence of the target class prediction results on the transfer of the nontarget class knowledge. This approach achieves efficient model compression. With 67% fewer parameters and only 1.2 G of computation, the model proposed in this paper still has a 87.6% success rate, while improving the model inference speed. [ABSTRACT FROM AUTHOR]

Deep learning has achieved excellent achievements and has become the mainstream in the field of aerial image classification. While obtaining remarkable success, deep learning-based approaches are notoriously dependent on large amounts of labelled data. Few-shot learning uses existing knowledge to learn from few samples and quickly generalizes to new tasks. In this work, we proposed the few-shot learning with deep economic network and teacher knowledge for aerial image classification. Firstly, we performed simplification twice to reduce large-scale parameters and computational effort in deep networks. In the first simplification, the redundancy in feature inputs' main components is reduced, and the implicit information in redundant components is extracted instead of directly discarding the redundant components. The channel and spatial redundancies in deeper layers' inputs are reduced in the second simplification. Secondly, the teacher knowledge guides the random sampling and uses limited samples to improve classification performance. We conducted extensive experiments on NWPU-RESISC45, RSD46-WHU, and UC Merced datasets. The experimental results reveal that the proposed method has better classification accuracy, fewer network parameters, and less computational effort. Experiments on miniImageNet, FC100, CUB, and cross-domain datasets show that our method also maintains advanced classification accuracy on few-shot image classification benchmarks. [ABSTRACT FROM AUTHOR]

In this article, we propose a novel framework for camouflaged object detection (COD), named D $^{2}$ C-Net, which contains two new modules: Dual-branch features extraction (DFE) and gradually refined cross fusion (GRCF). Specifically, the DFE simulates the two-stage detection process of human visual mechanisms in observing camouflage scenes. For the first stage, a dense concatenation is employed to aggregate multilevel features and expand the receptive field. The first stage feature maps are then utilized to extract two-direction guidance information, which benefits the second stage. The GRCF consists of a self-refine attention unit and a cross-refinement unit, with the aim of combining the peer layer features and DFE features for an improved COD performance. The proposed framework outperforms 13 state-of-the-art deep learning-based methods upon three public datasets in terms of five widely used metrics. Finally, we show evidence for the successful applications of the proposed method in the fields of surface defect detection and medical image segmentation. [ABSTRACT FROM AUTHOR]

Magnesium ions (Mg2+) are essential for the folding, functional expression, and structural stability of RNA molecules. However, predicting Mg2+–binding sites in RNA molecules based solely on RNA structures is still challenging. The molecular surface, characterized by a continuous shape with geometric and chemical properties, is important for RNA modelling and carries essential information for understanding the interactions between RNAs and Mg2+ ions. Here, we propose an approach named RNA–magnesium ion surface interaction fingerprinting (RMSIF), a geometric deep learning-based conceptual framework to predict magnesium ion binding sites in RNA structures. To evaluate the performance of RMSIF, we systematically enumerated decoy Mg2+ ions across a full-space grid within the range of 2 to 10 Å from the RNA molecule and made predictions accordingly. Visualization techniques were used to validate the prediction results and calculate success rates. Comparative assessments against state-of-the-art methods like MetalionRNA, MgNet, and Metal3DRNA revealed that RMSIF achieved superior success rates and accuracy in predicting Mg2+-binding sites. Additionally, in terms of the spatial distribution of Mg2+ ions within the RNA structures, a majority were situated in the deep grooves, while a minority occupied the shallow grooves. Collectively, the conceptual framework developed in this study holds promise for advancing insights into drug design, RNA co-transcriptional folding, and structure prediction. [ABSTRACT FROM AUTHOR]

• The deep learning pneumothorax segmentation method obtained state-of-the-art results. • The SDM regression method provided a solution to the blurred boundary problem. • A new pneumothorax segmentation dataset PTX-498 has been released. Accurate segmentation of the pneumothorax from chest X-ray images is of considerable significance for the diagnosis in the emergency department. Recently, deep learning methods have achieved impressive progress in many medical image segmentation tasks. However, the blurred boundary of pneumothorax in chest X-ray images makes it difficult for these algorithms to perform well in pneumothorax segmentation. In this paper, we aim to solve the problem in current learning-based segmentation methods that they often ignore the information of the target boundary and thus fail to generate the result with the correct shape. The proposed new learning framework, DeepSDM, employs the rich information in Signed Distance Map (SDM). The binary segmentation mask and SDM are learned in parallel via the multi-task strategy. Moreover, a boundary-based weighting scheme is adopted in the SDM regression task, forcing the model to focus more on pneumothorax and its contour. The proposed DeepSDM is validated by extensive experiments on the Kaggle SIIM-ACR Pneumothorax Segmentation dataset and our PTX-498 dataset. The results demonstrate that DeepSDM effectively perceives the boundary of pneumothorax and outperforms other state-of-the-art methods. To provoke the research interest in the community, the PTX-498 dataset, codes, and trained models are publicly available at Zenodo and GitHub. [ABSTRACT FROM AUTHOR]

The main purpose of image Co-saliency detection is to detect objects with similar properties or prospects in a group of images. In this paper, we propose a novel Converged Complementary Co-saliency detection framework in an end-to-end manner for Co-saliency detection. Specifically, considering the consistency and the difference between the features of the intra-image and the inter-image, our network contains two parts [forward information transfer module and complementary information enhancement module (CIEM)]. Each part is a double-branch network with short connections to avoid incomplete information transmission, including single-saliency detection branch (SSDB) and Co-saliency detection branch (CSDB). In FITM, the SSDB and CSDB extract the corresponding features, respectively. In order to produce more accurate results, we feed the SSDB and CSDB into the holistic attention and perform further feature extraction. In CIEM, we extract the useful information of the non-salient area. At the end of the network, we fuse the FITM and CIEM to generate the final Co-saliency map. Furthermore, we rearranged a new data set based on the existing public data set. Compared with other Co-saliency methods, the experimental results show that our proposed model achieves state-of-the-art performance. [ABSTRACT FROM AUTHOR]

Camouflaged object detection (COD) is an emerging visual detection task, which aims to locate and distinguish the disguised target in complex backgrounds by imitating the human visual detection system. Recently, COD has attracted increasing attention in computer vision, and a few models of camouflaged object detection have been successfully explored. However, most existing works primarily focus on modeling camouflaged object detection over in-depth analyzing existing COD structures. To the best of our knowledge, a systematic review for COD has not been publicly reported, especially for recently proposed deep learning-based COD models. To make up this vacancy, we firstly proposed a comprehensive review on both COD models and public benchmark datasets and provide potential directions for future COD studies. Specifically, we conduct a comprehensive summary of 39 existing COD models from 1998 to 2021. And then, to facilitate subsequent research on COD, we classify the existing structures into two categories, 27 traditional handcrafted feature-based structures and 12 structures based on deep learning. In addition, we further group traditional handcrafted feature-based structures into six sub-classes based on the detection mechanism: texture, color, motion, intensity, optical flow, and multi-modal fusion. Furthermore, we take an in-depth analysis of the deep learning-based structure based on both detection motivation and detection performance and evaluate the performance of each structure. Moreover, we sum up four widely used COD datasets and describe the details of each one. Finally, we also discuss the limitations of COD and the corresponding solutions to improve detection accuracy. We still mention the relevant applications of camouflaged object detection and its future research directions to promote the development of camouflaged object detection. [ABSTRACT FROM AUTHOR]

Simple Summary: The timely and accurate detection of ewe estrus behavior in precision animal husbandry is an important research topic. The timely detection of estrus ewes in mutton sheep breeding can not only protect the welfare of ewes themselves, but also improve the interests of breeding enterprises. With the continuous increase in the scale of mutton sheep breeding and the gradual intensification of breeding methods, the traditional manual detection methods require high labor intensity, and the contact sensor detection methods will cause stress reaction problems in ewes. In recent years, the rapid development of deep learning technology has brought on new possibilities. We propose a method for the recognition ewe estrus based on a multi-target detection layer neural network. The results show that the method can meet the requirements of timely and accurate detection of ewe estrus behavior in large-scale mutton sheep breeding. There are some problems with estrus detection in ewes in large-scale meat sheep farming: mainly, the manual detection method is labor-intensive and the contact sensor detection method causes stress reactions in ewes. To solve the abovementioned problems, we proposed a multi-objective detection layer neural network-based method for ewe estrus crawling behavior recognition. The approach we proposed has four main parts. Firstly, to address the problem of mismatch between our constructed ewe estrus dataset and the YOLO v3 anchor box size, we propose to obtain a new anchor box size by clustering the ewe estrus dataset using the K-means++ algorithm. Secondly, to address the problem of low model recognition precision caused by small imaging of distant ewes in the dataset, we added a 104 × 104 target detection layer, making the total target detection layer reach four layers, strengthening the model's ability to learn shallow information and improving the model's ability to detect small targets. Then, we added residual units to the residual structure of the model, so that the deep feature information of the model is not easily lost and further fused with the shallow feature information to speed up the training of the model. Finally, we maintain the aspect ratio of the images in the data-loading module of the model to reduce the distortion of the image information and increase the precision of the model. The experimental results show that our proposed model has 98.56% recognition precision, while recall was 98.04%, F1 value was 98%, mAP was 99.78%, FPS was 41 f/s, and model size was 276 M, which can meet the accurate and real-time recognition of ewe estrus behavior in large-scale meat sheep farming. [ABSTRACT FROM AUTHOR]

Background: Due to the high cost of data collection for magnetization detection of media, the sample size is limited, it is not suitable to use deep learning method to predict its change trend. The prediction of physical and chemical properties of magnetized water and fertilizer (PCPMWF) by meta-learning can help to explore the effects of magnetized water and fertilizer irrigation on crops. Method: In this article, we propose a meta-learning optimization model based on the meta-learner LSTM in the field of regression prediction of PCPMWF. In meta-learning, LSTM is used to replace MAML's gradient descent optimizer for regression tasks, enables the meta-learner to learn the update rules of the LSTM, and apply it to update the parameters of the model. The proposed method is compared with the experimental results of MAML and LSTM to verify the feasibility and correctness. Results: The average absolute percentage error of the meta-learning optimization model of meta-learner LSTM is reduced by 0.37% compared with the MAML model, and by 4.16% compared with the LSTM model. The loss value of the meta-learning optimization model in the iterative process drops the fastest and steadily compared to the MAML model and the LSTM model. In cross-domain experiments, the average accuracy of the meta-learning optimized model can still reach 0.833. Conclusions: In the case of few sample, the proposed model is superior to the traditional LSTM model and the basic MAML model. And in the training of cross-domain datasets, this model performs best. [ABSTRACT FROM AUTHOR]

Wind power prediction is crucial for successfully integrating large-scale wind energy with the grid and achieving a carbon-neutral energy mix. However, previous studies encountered challenges to noise reduction, information extraction, error handling, and uncertainty estimation. To address these limitations, this study proposed a novel wind power forecasting system, systematically enhancing forecasting capability by refining the raw time series using horizontal denoising and vertical granulation methods, utilizing a bidirectional deep learning model for capturing nonlinear features and optimizing its core hyperparameters with a newly developed multi-objective enhanced version of metaheuristic algorithm to achieve Pareto optimal solutions, employing the traditional statistical approach for residual linear feature extraction from error series and effective correction, and integrating quantile regression for interval prediction. The predictive performance was evaluated based on wind power, speed, and direction data from two different countries, the system yielded correlation coefficients of 0.9796, 0.9943, 0.9795, and 0.9780 in spring, summer, autumn and winter for China, and 0.9950 for Turkey, demonstrating its pronounced advantages over comparative models. Overall, the proposed system contributes to the advancement of wind-power prediction theory and holds substantial potential for practical applications. • A novel hybrid wind power forecasting system was proposed. • Solve problems of noise interference and high computing burden by data refining. • A strategy to fully extract nonlinear and linear data features was designed. • An enhanced multi-objective golden jackal algorithm was raised for optimization. • Pareto optimal solutions were theoretically proven to be achieved. [ABSTRACT FROM AUTHOR]

Hazy weather seriously threatens human health, providing accurate and reliable air pollutant concentration forecast will be of great scientific value and benefit to residents' livelihood and natural environment. However, existing forecasting models lack outlier detection, feature mapping and uncertainty forecast, which results in unreliable forecasting results. Thus, in this paper, a novel combined deterministic and probabilistic forecasting system based on meta -heuristic algorithm is proposed, which contains multiple data pre-processing, combined deterministic and probabilistic forecast. Firstly, the raw data is pre-processed and analyzed from multiple perspectives to fully extract the chaotic features. Then multi-objective white shark optimizer is utilized to combine the three deep learning models to enhance the forecasting accuracy and stability. Furthermore, the Pin-ball loss function is employed to yield prediction intervals at different confidence levels, and a novel interval pseudo-code is constructed to optimize the prediction intervals thus promoting the probabilistic forecasting reliability. The simulations indicated that the developed system enhances the forecasting accuracy by maximum 60.6772%, 54.9793% and 38.6876% in the three steps of PM 2.5 prediction and up to 64.9001%, 68.4637% in the interval sharpness at significance level 0.05 and 0.1 than the tested models. [ABSTRACT FROM AUTHOR]

• A novel feature processing paradigm and the model based on it are proposed. • The rationality of the method is corroborated in hydrological feature discovery and time-series dependence extraction. • The proposed method shows its excellent performance in term of runoff prediction. • The performance of all models is validated on two proposed constructing noisy datasets. The interaction between hydrological factors is complex and the correlation effects cannot be quantitatively explained from a mechanistic perspective. The extraction of effective features from several hydrological and meteorological data and quantifying their correlation effects to make runoff prediction more accurate and stable is an urgent problem to be solved. In this study, we introduce a structural paradigm for hydrological time series to handle the entire feature space composed of some feature sequences based on deep learning methods, namely, the feature attention mechanism. This method transforms the problem of feature-target association into multiple parallel binary classification problems and assigns attention units to each specific feature in the network. The attention distribution was adjusted by updating the neural network parameters in a supervised manner to generate feature weightings. In addition, two extensions were proposed based on the initial network structure. The three methods were used in a practical study of the upper Yangtze River Basin using three evaluation metrics compared with five benchmark methods. The mean metric values were improved by up to 9.43, 10.05, and 2.65%, respectively. Meanwhile, with the characteristics of hydrological data, the rationality of the method is corroborated in hydrological feature discovery and time-series dependence extraction. Moreover, we tested the performance of the model on two constructed noise datasets to demonstrate its robustness. The results of all the above experiments show that the attention module significantly improves the learning and generalization ability, enhances noise resistance, and strengthens the robustness of the model compared with the traditional method. [ABSTRACT FROM AUTHOR]