Adequate Data Sets Research Articles

Using Synthetic Images to Evaluate and Improve Object Detection Neural Network Performance on Aerial Image Datasets

Machine learning-based models for object detection heavily rely on large datasets of labeled images. When models trained on these datasets are applied to Unmanned Aerial Vehicle (UAV) imagery, the problem arises that the conditions under which the training images were created (lighting, altitude, angle) may be different to the UAVs applied conditions, leading to misclassifications. This problem becomes even more pressing in safety critical applications where failures can have huge negative impacts and constitute obstacles for certification of cognitive UAV components. Along a case study on car detection in low-altitude aerial imagery, we show that using, both, artificial and real images for model training has a positive effect on the performance of object detection algorithms when the trained model is applied on images from another domain. Additionally, we show that weak points from object detection neural networks trained on real-world images transfer to synthetic images and that synthetic data can be used to evaluate neural networks trained on real-world data. Since simulated images are easy to create and object labels are inherently given, the presented approaches show a promising direction for scenarios where adequate datasets are difficult to obtain, as well as for the targeted exploration of weak points of object detection algorithms.

Towards lifelong object recognition: A dataset and benchmark

Lifelong learning algorithms aim to enable robots to handle open-set and detrimental conditions, and yet there is a lack of adequate datasets with diverse factors for benchmarking. In this work, we constructed and released a lifelong learning robotic vision dataset, OpenLORIS-Object. This dataset was collected by RGB-D camera capturing dynamic environment in daily life scenarios with diverse factors, including illumination, occlusion, object pixel size and clutter, of quantified difficulty levels. To the best of our knowledge, this is an unique real-world dataset for robotic vision with independent and quantifiable environmental factors, which are currently unaccounted for in other lifelong learning datasets such as CORe50 and NICO. We tested 9 state-of-the-art algorithms with 4 evaluation metrics over the dataset in Domain Incremental Learning, Task Incremental Learning, and Class Incremental Learning scenarios. The results demonstrate that these existing algorithms are insufficient to handle lifelong learning task in dynamic environments. Our dataset and benchmarks are now publicly available at this website.22https://lifelong-robotic-vision.github.io/dataset/object

Spatial indicators as a tool to support the decision-making process in relation to different goals of rural planning

Indicators based on the application of Geographic Information System (GIS) techniques are widely used at all levels of spatial policy. The bulk of studies undertaken have involved examinations of urban and peri-urban areas, but only a few of them have been developed in relation to rural areas. To address this gap, the current paper aims to analyse Spatial Planning Indicators (SPI) based on the data obtained from spatial databases with reference to four scales: rural municipalities, rural villages, public rural spaces and singular building plot from the point of view of supporting the decision-making process. A review of the relevant literature (102 papers) enabled the selection of a set of indices applied for four specific purposes of spatial planning including: (1) landscape/ecological values; (2) sustainability development; (3) risk assessment; and (4) spatial dimension of urbanisation. Then, SPI were calculated in relation to four spatial scales of Spiczyn municipality (Poland). The effectiveness analysis was conducted based on five criteria related to relevance, measurability and applicability. Generally, the analysis demonstrated that the effectiveness of different kinds of indices depends strongly on each specific field of application, and they were found to be more relevant in large geographic regions than smaller ones. These findings indicate the need for the careful selection of SPI to obtain an adequate dataset that is useful for the management of rural areas. Each spatial scale required separate consideration when drawing conclusions with regard to land use planning and the elaboration of relevant indices.

ZeVigilante: Detecting Zero-Day Malware Using Machine Learning and Sandboxing Analysis Techniques.

For the enormous growth and the hysterical impact of undocumented malicious software, otherwise known as Zero-Day malware, specialized practices were joined to implement systems capable of detecting these kinds of software to avert possible disastrous consequences. Owing to the nature of developed Zero-Day malware, distinct evasion tactics are used to remain stealth. Hence, there is a need for advance investigations of the methods that can identify such kind of malware. Machine learning (ML) is among the promising techniques for such type of predictions, while the sandbox provides a safe environment for such experiments. After thorough literature review, carefully chosen ML techniques are proposed for the malware detection, under Cuckoo sandboxing (CS) environment. The proposed system is coined as Zero-Day Vigilante (ZeVigilante) to detect the malware considering both static and dynamic analyses. We used adequate datasets for both analyses incorporating sufficient samples in contrast to other studies. Consequently, the processed datasets are used to train and test several ML classifiers including Random Forest (RF), Neural Networks (NN), Decision Tree (DT), k-Nearest Neighbor (kNN), Naïve Bayes (NB), and Support Vector Machine (SVM). It is observed that RF achieved the best accuracy for both static and dynamic analyses, 98.21% and 98.92%, respectively.

Deep Neural Network-Based Removal of a Decaying DC Offset in Less Than One Cycle for Digital Relaying

To make a correct decision during normal and transient states, the signal processing for relay protection must be completed and designated the correct task within the shortest given duration. This paper proposes to solve a dc offset fault current phasor with harmonics and noise based on a Deep Neural Network (DNN) autoencoder stack. The size of the data window was reduced to less than one cycle to ensure that the correct offset is rapidly computed. The effects of different numbers of the data samples per cycle are discussed. The simulations revealed that the DNN autoencoder stack reduced the size of the data window to approximately 90% of a cycle waveform, and that DNN performance accuracy depended on the number of samples per cycle (32, 64, or 128) and the training dataset used. The fewer the samples per cycle of the training dataset, the more training was required. After training using an adequate dataset, the delay in the correct magnitude prediction was better than that of the partial sums (PSs) method without an additional filter. Similarly, the proposed DNN outperformed the DNN-based full decay cycle dc offset in the case of converging time. Taking advantage of the smaller DNN size and rapid converging time, the proposed DNN could be launched for real-time relay protection and centralized backup protection.

Real-time goat face recognition using convolutional neural network

Automatic identification of individual animals is an important step towards achieving accurate breeding histories, significant contributions to breeding and genetic management programmers. Currently, a different type of tags, tattoos, paint brands and microchips are used to uniquely identify livestock animals. However, the manual identification system is time-consuming, expensive and unreliable. In this paper, we present a deep learning approach that aims to fully automated pipeline for face detection and recognition of goats. Due to the high similarity and the lack of adequate dataset this problem is more complex than human face recognition. We composed two different publicly available datasets for detection and recognition. State-of-the-art convolutional neural networks (CNN) model are trained on this dataset. To evaluate the robustness of our approach, we compared it with different face recognition methods. The results show better performance with an accuracy of 96.4%. Furthermore, this paper reports 93%, 83%, 92% and 85% detection accuracy (average precision) for face, eye, nose and ear, respectively. The findings of this research could be helpful to improve animal health and welfare, individual monitoring, activity monitoring and phenotypic data collection. All the dataset and the related outcome are publicly available (https://doi.org/10.17632/4skwhnrscr.2).

Graph Neural Networks as a Potential Tool in Improving Virtual Screening Programs.

Despite the increasing number of pharmaceutical companies, university laboratories and funding, less than one percent of initially researched drugs enter the commercial market. In this context, virtual screening (VS) has gained much attention due to several advantages, including timesaving, reduced reagent and consumable costs and the performance of selective analyses regarding the affinity between test molecules and pharmacological targets. Currently, VS is based mainly on algorithms that apply physical and chemistry principles and quantum mechanics to estimate molecule affinities and conformations, among others. Nevertheless, VS has not reached the expected results concerning the improvement of market-approved drugs, comprising less than twenty drugs that have reached this goal to date. In this context, graph neural networks (GNN), a recent deep-learning subtype, may comprise a powerful tool to improve VS results concerning natural products that may be used both simultaneously with standard algorithms or isolated. This review discusses the pros and cons of GNN applied to VS and the future perspectives of this learnable algorithm, which may revolutionize drug discovery if certain obstacles concerning spatial coordinates and adequate datasets, among others, can be overcome.

Growth of Simulated Tumors Under the Influence of Oxygen Supply

Utilizing machine learning techniques for classification of tumorous tissue has been shown to be extremely effective for external imaging and processing. However, when training an algorithm to work intraoperatively for identifying the extent of cancerous tissue within the body using intraoperative sensors, an adequate data set is not readily available or in many cases not feasible to obtain. For this reason, accurate and realistic simulations of cancer growth can be a powerful tool in supplementing datasets where it is not realistic to directly obtain sensor measurements in the body. This work introduces a simulation environment to generate cell-based cancerous tumors grown within a controlled oxygen environment. The growth and oxygen simulations are outlined along with how they are implemented within an open source simulation environment. Simulation results are provided showing the influence of the oxygen field on tumor growth shape and rate. The results follow intuitive growth characteristics and exhibit invasive behavior from lack of oxygen. The simulation environment lends itself to be easily and flexibly used for further detailed investigations of tumor growth.

Building and Testing Fine-Grained Dataset of COVID-19 Tweets for Worry Prediction

The COVID-19 outbreak has resulted in the loss of human life worldwide and has increased worry concerning life, public health, the economy, and the future. With lockdown and social distancing measures in place, people turned to social media such as Twitter to share their feelings and concerns about the pandemic. Several studies have focused on analyzing Twitter users’ sentiments and emotions. However, little work has focused on worry detection at a fine-grained level due to the lack of adequate datasets. Worry emotion is associated with notions such as anxiety, fear, and nervousness. In this study, we built a dataset for worry emotion classification called “WorryCov” . It is a relatively large dataset derived from Twitter concerning worry about COVID-19. The data were annotated into three levels (“no-worry”, “worry”, and “high-worry”). Using the annotated dataset, we investigated the performance of different machine learning algorithms (ML), including multinomial Naïve Bayes (MNB), support vector machine (SVM), logistic regression (LR), and random forests (RF). The results show that LR was the optimal approach, with an accuracy of 75%. Furthermore, the results indicate that the proposed model could be used by psychologists and researchers to predict Twitter users’ worry levels during COVID-19 or similar crises.

Toward A Holistic, Efficient, Stacking Ensemble Intrusion Detection System using a Real Cloud-based Dataset

Network intrusion detection is a key step in securing today’s constantly developing networks. Various experiments have been put forward to propose new methods for resisting harmful cyber behaviors. Though, as cyber-attacks turn out to be more complex, the present methodologies fail to adequately solve the problem. Thus, network intrusion detection is now a significant decision-making challenge that requires an effective and intelligent approach. Various machine learning algorithms such as decision trees, neural networks, K nearest neighbor, logistic regression, support vector machine, and Naive Bayes have been utilized to detect anomalies in network traffic. However, such algorithms require adequate datasets to train and evaluate anomaly-based network intrusion detection systems. This paper presents a testbed that could be a model for building real-world datasets, as well as a newly generated dataset, derived from real network traffic, for intrusion detection. To utilize this real dataset, the paper also presents an ensemble intrusion detection model using a meta-classification approach enabled by stacked generalization to address the issue of detection accuracy and false alarm rate in intrusion detection systems.

Classifying Early Blight and Late Blight of Potato Based on Convolution Neural Network

HighlightsA convolution neural network (CNN) model, called M_Net, was designed to recognize early blight and late blight in potato.Our model achieves the highest accuracy with low computation requirements compare with other popular deep neural networks.Hyperparameters tuning is performed to optimize accuracy, generalization, and computation requirements for potato disease classification.Experimental results show that the combination of multiple datasets improves the generalization of the model.Abstract.Early and late blight are two of the most common potato diseases. Intelligent tools for automatically detecting these two diseases can benefit farmers and agricultural extension officers. However, it remains a challenge to use traditional image processing methods to recognize these diseases. Convolution neural network (CNN) is an advanced methodology in computer vision, which shows great promise in image classification. This article explores CNN models to classify potato early blight and late blight based on leaf images. This research task has three challenges: lack of adequate datasets, noise in existing data, and the construction of a model that handles variability in image backgrounds. This research designs a CNN model M_Net based on MobileNetV1 network and uses different dataset sources in the construction of a CNN model with a strong generalization ability to identify disease leaves and healthy leaves. Furthermore, this article adds a new dataset to the field by supplying the model with potato leaf images. The results show that the CNN model achieves the highest accuracy with low calculation cost compared to some classical models and the final model has a strong generalization capacity. Keywords: Accuracy, CNN, Early blight, Generalization ability, Late blight.

LSDStrategy: A Lightweight Software-Driven Strategy for Addressing Big Data Variety of Multimedia Streaming

Many devices, users, and applications stream an irregular amount of varied data every second. This rapid generation of data continues at an enormous rate, constructing the big data that increase the need for solutions, despite resource constraints, to analyze and manipulate data. Current methods allocate cloud resources according to the characteristics of the data. Resource allocation requires a comprehensive view of the workload requirements. However, the data characteristics in big data streams are uncertain due to the random nature of data generation. Choosing and allocating the right resources to this stream is challenging. With the variety of big data streams, the stochastic nature of the stream led to unpredictable requirements and specifications. The critical issue is forecasting the workload to avoid the over-provisioning and under-provisioning of resources. Such forecasting needs an adequate dataset to describe the history logs of the incoming workload. A fast release for such a dataset provides a high chance of deploying forecasting at the right time. This paper addresses this issue with a novel strategy named LSDStrategy that analyzes the received multimedia stream based on its binary content using machine learning techniques with artificial and real datasets. LSDStrategy utilizes an evaluating voting technique to select the optimum classifier to trade off accuracy and prediction time as metrics. Multi-classifiers that have been built and tested include Decision Tree (DT), K-Nearest Neighbor (K-NN), and Random Forest (RF) over multi-content-based features. Experiments evaluated the performance of the adopted models and the selected features. According to experimental analysis, the DT approach provides a consistent performance for both artificial and real-world datasets for 85% and 81.3%, respectively. We deploy and evaluate the LSDStrategy efficiency on a regular specification server through a set of experiments using a synthetic stream. The experiments prove the LSDStrategy agility and adaptivity in identifying the multimedia-based workload type utilizing small chunks of load.

Classifying news articles in multiple languages: leveraging context aware models

Despite the recent advances in text classification and the performance improvement yielded by Transformers models, the absence or inaccessibility of an adequate dataset to train a text classifier motivates the choice for alternative routes. In this study, the need to detect specific topics in the news and to discard irrelevant content encouraged the development of an article tagging pipeline which assesses the similarity between a user-defined dictionary of topic-specific keywords and news article keywords. The innovation of the paper stands in the exploitation of two BERT-based algorithms to retrieve article keywords and to embed them, which previous studies have shown to outperform state of the art solutions for keywords extraction and semantic textual similarity. In a nutshell, the pipeline computes the semantic similarity between the sentence embeddings generated from topic-specific keywords and those produced from news article keywords extracted with the KeyBERT algorithm, finally classifying each article according to a previously defined topic. The results are supported by sound coherence and diversity metrics computed, by aggregating each article by their first tag, which attests to the semantic validity of the pipeline outputs.

View VULMA: Data Set for Training a Machine-Learning Tool for a Fast Vulnerability Analysis of Existing Buildings

The paper presents View VULMA, a data set specifically designed for training machine-learning tools for elaborating fast vulnerability analysis of existing buildings. Such tools require supervised training via an extensive set of building imagery, for which several typological parameters should be defined, with a proper label assigned to each sample on a per-parameter basis. Thus, it is clear how defining an adequate training data set plays a key role, and several aspects should be considered, such as data availability, preprocessing, augmentation and balancing according to the selected labels. In this paper, we highlight all these issues, describing the pursued strategies to elaborate a reliable data set. In particular, a detailed description of both requirements (e.g., scale and resolution of images, evaluation parameters and data heterogeneity) and the steps followed to define View VULMA are provided, starting from the data assessment (which allowed to reduce the initial sample of about 20.000 images to a subset of about 3.000 pictures), to achieve the goal of training a transfer-learning-based automated tool for fast estimation of the vulnerability of existing buildings from single pictures.

Machine Learning-Based Lie Detector Applied to a Novel Annotated Game Dataset

Lie detection is considered a concern for everyone in their day-to-day life, given its impact on human interactions. Thus, people normally pay attention to both what their interlocutors are saying and to their visual appearance, including the face, to find any signs that indicate whether or not the person is telling the truth. While automatic lie detection may help us to understand these lying characteristics, current systems are still fairly limited, partly due to lack of adequate datasets to evaluate their performance in realistic scenarios. In this work, we collect an annotated dataset of facial images, comprising both 2D and 3D information of several participants during a card game that encourages players to lie. Using our collected dataset, we evaluate several types of machine learning-based lie detectors in terms of their generalization, in person-specific and cross-application experiments. We first extract both handcrafted and deep learning-based features as relevant visual inputs, then pass them into multiple types of classifier to predict respective lie/non-lie labels. Subsequently, we use several metrics to judge the models’ accuracy based on the models predictions and ground truth. In our experiment, we show that models based on deep learning achieve the highest accuracy, reaching up to 57% for the generalization task and 63% when applied to detect the lie to a single participant. We further highlight the limitation of the deep learning-based lie detector when dealing with cross-application lie detection tasks. Finally, this analysis along the proposed datasets would potentially be useful not only from the perspective of computational systems perspective (e.g., improving current automatic lie prediction accuracy), but also for other relevant application fields, such as health practitioners in general medical counselings, education in academic settings or finance in the banking sector, where close inspections and understandings of the actual intentions of individuals can be very important.

Normal Values of Left Ventricular Size and Function on Three-Dimensional Echocardiography: Results of the World Alliance Societies of Echocardiography Study

Echocardiography remains the most widely used modality to assess left ventricular (LV) chamber size and function. Currently this assessment is most frequently performed using two-dimensional (2D) echocardiography. However, three-dimensional (3D) echocardiography has been shown to be more accurate and reproducible than 2D echocardiography. Current normative reference values for 3D LV analysis are based predominantly on data from North America and Europe. The World Alliance Societies of Echocardiography study was designed to sample normal subjects from around the world to provide more universal global reference ranges. The aim of this study was to assess the worldwide feasibility of LV 3D echocardiography and report on size and functional measurements. A total of 2,262 healthy subjects were prospectively enrolled from 19 centers in 15 countries. Three-dimensional LV full-volume data sets were obtained and analyzed offline using vendor-neutral software. Measurements included LV end-diastolic and end-systolic volumes, LV ejection fraction (LVEF), global longitudinal strain (GLS), and global circumferential strain. Results were categorized by age (18-40, 41-65, and >65years), sex, and race. A total of 1,589 subjects (feasibility 70%) had adequate LV data sets for analysis. Mean normal values for indexed end-diastolic volume, end-systolic volume, and LVEF in men and women were 70±15 and 65±12mL/m2, 28±7 and 25±6mL/m2, and 60±5% and 62±5%, respectively. Men had larger LV volumes and lower LVEFs than women. GLS and global circumferential strain were higher in magnitude in women. In both sexes, LV volumes were lower and LVEF tended to be higher with increasing age, especially considering the differences between the youngest and oldest age groups. Although GLS was similar across age groups in men, in women, the youngest and middle-age cohorts revealed higher magnitudes of GLS compared with the oldest age group. Global circumferential strain was higher in magnitude at older age in both men and women. Finally, Asians had smaller chamber sizes and higher LVEFs and absolute strain values than both blacks and whites. Age, sex, and race should be considered when defining normal reference values for LV dimension and functional parameters obtained by 3D echocardiography.

Honey bees as biomonitors – Variability in the elemental composition of individual bees

Homogenized or pooled samples of honey bees are already used for monitoring of metal pollution in the environment for a couple of decades. This is the first study that analyzed the elemental composition of individual honey bees. One apiary with 21 hives was used to test in-hive and between-hive differences at the same location. Highest in-hive variability was observed for Al (14.1x), Li (10.3x), V (10.3x), As (8.9x) and Cd (7.9x). For the elements Cu, K, Mg, Na, P, S and Zn the smallest both in-hive (2.3x, 1.8x, 2.0x, 2.2x, 1.9x, 1.7x, 2.4x respectively) as well as between-hive variability (1.5x, 1.4x, 1.4x, 1.6x, 1.4x, 1.4x and 1.5x respectively) was observed. Nonetheless, between-hive variability was statistically significant for all the analyzed elements. We proved that these differences significantly influence comparison of element concentrations in bees from different locations. One hive from Mesić apiary had significantly lower concentrations of Al (hive 18: 18.6 mg kg−1) and U (hive 1: 0.0013 mg kg−1), while at the same time a different hive from the same apiary showed higher concentration of these elements (Al, hive 14: 125 mg kg−1; U, hive 13: 0.012 mg kg−1) compared to another location (thermal power plant, Al: 97 mg kg−1, U: 0.0044 mg kg−1). Therefore, when using honey bees as biomonitors, we recommend a larger number of hives to be sampled at each location to provide an adequate dataset for reliable interpretation of results. Comparing individual bee elemental concentrations, for Na, Mg, P, S, K, Fe, Cu, and Zn negative correlations were found between dry mass of a bee and concentrations of these elements. These negative correlations could be a consequence of higher honey or nectar content in some of the sampled bees. However, this should also be considered when using honey bees as bioindicators.

DNNAttention: A deep neural network and attention based architecture for cross project defect number prediction

Software defect prediction (SDP) is the process of detecting fault-prone classes or modules in a software system. It helps in allocating resources before the testing phase more optimally. Due to a lack of an adequate dataset, defects can be predicted by employing data from different projects to train the classifier called cross-project defect prediction (CPDP). Cross-project defect number prediction (CPDNP) is one step ahead of CPDP, in which we can also estimate the number of defects in each module of a software system; we contemplate it as a regression problem. This article dealt with the CPDNP mechanism and suggested a CPDNP architecture by employing a deep neural network and attention layer called DNNAttention. We syntheses substantial data named cross-heap by utilizing an amalgamation of 44 projects from the PROMISE repository. We fed the cross-heap into DNNAttention to train and evaluate the performance over 44 datasets by applying transfer learning. We have also address class imbalance (CI) and overfitting problems by employing multi-label random over-sampling and dropout regularization, respectively. We compared the performance of the DNNAttention using mean squared error (MSE), mean absolute error (MAE), and accuracy over eight baseline methods. We found out of 44 projects, 19 and 20 have minimum MSE and MAE, respectively, and in 19 projects, accuracy yields by the proposed model surpasses exiting techniques. We also compared the performance in terms of Kendall and Fault-Percentile-Average with the recent unsupervised method and found DNNAttention significantly outperforms this method. Moreover, we found the improvement of the DNNAttention over other baseline methods in terms of MAE, MSE, and accuracy by inspecting 20% line of code are substantial. In most situations, the improvements are significant, and they have a large effect size across all 44 projects.

Commensal small mammal trapping data in Southern Senegal, 2012–2015: where invasive species meet native ones

Describing patterns and testing hypotheses on processes driving biological invasions represent major issues in ecology. Addressing these questions requires building adequate data sets, i.e., covering areas and spanning periods adapted to the invasion processes studied. Rodents include major invasive species, among which the black rat Rattus rattus and the domestic mouse Mus musculus have nearly colonized the entire world, from their native Asian range. To do so, they have benefitted from their ability to cope with human-modified environments and to live in the immediate vicinity of Man, who served as a vector of their dispersal between regions and continents. In Senegal, both R.rattus and M.musculus, initially introduced by early West European colonizers some centuries ago, are currently expanding thanks to road traffic and infrastructure development and rampant urbanization that concerns even remote regions of the country. As part of projects aimed at studying (1) the role of invasive black rat populations in the emergence of zoonotic diseases in southeastern Senegal, and (2) the evolutionary consequences of parasites in R.rattus and M.musculus invasions in Senegal, we conducted a series of field campaigns throughout the southern half of the country, between May 2012 and September 2015. The objectives were to catch commensal small mammals using standard trapping procedures, identify them using morphological or molecular tools, and take samples from them upon autopsy, to look for zoonotic parasites and pathogens. Along with data on individual specimens, information on microhabitats was gathered at each trap position. This resulted in the constitution of a data set of more than 13,000 trapnights, which allowed the capture of more than 3,100 small mammals, all characterized by a series of associated biological, geographical, and environmental data. The small mammals concerned are mainly rodents (10 species), shrews, and hedgehogs. The two invasive rodent species were the most numerous, exceeding in numbers all the other species pooled. This data set makes it possible to study coarse to fine-scaled distribution of species of this commensal community in southern Senegal, as well as the possible determinants of this distribution in terms of habitat preferences and/or interspecific interactions. This data set can be freely used for non-commercial purposes and is licensed under a Creative Commons Attribution 4.0 International License.

Underwater Acoustic Target Recognition Based on Generative Adversarial Network Data Augmentation

Underwater acoustic target recognition is an important aspect of underwater acoustic research. In recent years, machine learning has been developed continuously, which is widely and effectively applied in underwater acoustic target recognition. In order to acquire good recognition results and reduce the problem of overfitting, Adequate data sets are essential. However, underwater acoustic samples are relatively rare, which has a certain impact on recognition accuracy. In this paper, in addition of the traditional audio data augmentation method, a new method of data augmentation using generative adversarial network is proposed, which uses generator and discriminator to learn the characteristics of underwater acoustic samples, so as to generate reliable underwater acoustic signals to expand the training data set. The expanded data set is input into the deep neural network, and the transfer learning method is applied to further reduce the impact caused by small samples by fixing part of the pre-trained parameters. The experimental results show that the recognition result of this method is better than the general underwater acoustic recognition method, and the effectiveness of this method is verified.