Self-made Dataset Research Articles

Artificial intelligence–based method for the rapid detection of fish parasites (Ichthyophthirius multifiliis, Gyrodactylus kobayashii, and Argulus japonicus)

Ichthyophthirius (Ichthyophthirius multifiliis), Monogenea (Gyrodactylus kobayashii) and fish lice (Argulus japonicus) are mainly infectious parasites, representative species of Protozoa, Platyhelminthes and Arthropoda, which cause serious economic losses in aquatic industry. In this research, a visual system that can rapidly detect and count these three kinds of parasites was realized based on a one-stage object detection deep learning algorithm YOLOv4 through python. Firstly, we made a dataset of parasites containing 27,930 images. Secondly, weights of the trained fish lice model were applied as the pre-training weights, and network (backbone indeed) frozen was also applied to obtaining a good performance predicting model with less time and higher accuracy, which showed that Transfer Learning could meet the training requirement for detecting these three fish parasites by using self-made data set. In addition, by comparison of different one-stage algorithms YOLOv4‑tiny, YOLOv3, et al., the best model with a total average accuracy (mAP) of 95.41% was achieved by the YOLOv4. Finally, this model could quickly detect and count mixed infected pictures with a speed of 0.13 s per image measured in GPU time. Further, a visual prediction and counting system equipped with the YOLOv4 was developed by using PyQt which is convenient for real-time video detection. A simple drug-giving system equipped with Praziquantel was also developed based on the thought of the Internet of Things in this study and after using a drug, the number of monogeneans infecting gold fish was reduced. At the same time, we modified YOLOv4 PANet by adding additional detection layers, which achieved greater performance of detecting smaller targets like Monogenea. Together, this Artificial intelligence–based method could realize the rapid detection and diagnosis of fish parasites in images and video.

TOF-UNet: High-precision method for terry towel defect detection

Towel defect detection mostly relies on manual labor, but there are problems such as a low efficiency and high missed detection rate. Therefore, automatic detection of towel defects is becoming increasingly popular. Although the UNet-based method has been successful, there are problems that must be solved for practical applications. To address the problems of the complex background caused by loops on the towel surface, relatively small defect size, and imbalanced defect–background ratio, a high-precision convolutional neural network is proposed, which is called tiny object-focused UNet. A coordinate attention mechanism is introduced in tiny object-focused UNet to enhance the feature-extraction capabilities, and spatial pyramid pooling is employed to fuse local and global information for more accurately extraction of towel defect features. Finally, the composite loss function obtained via the addition of the cross-entropy loss and the Dice loss function is used to reduce the impact of the imbalance in the proportion of defects on the detection accuracy. The proposed model is evaluated using a self-made dataset. The experimental results indicate that the segmentation performance of the network is better than that of other networks. Thus, the proposed method is useful for segmenting towel defects.

Fall Detection for Shipboard Seafarers Based on Optimized BlazePose and LSTM.

Aiming to avoid personal injury caused by the failure of timely medical assistance following a fall by seafarer members working on ships, research on the detection of seafarer’s falls and timely warnings to safety officers can reduce the loss and severe consequences of falls to seafarers. To improve the detection accuracy and real-time performance of the seafarer fall detection algorithm, a seafarer fall detection algorithm based on BlazePose–LSTM is proposed. This algorithm can automatically extract the human body key point information from the video image obtained by the vision sensor, analyze its internal data correlation characteristics, and realize the process from RGB camera image processing to seafarer fall detection. This fall detection algorithm extracts the human body key point information through the optimized BlazePose human body key point information extraction network. In this section, a new method for human bounding-box acquisition is proposed. In this study, a head detector based on the Vitruvian theory was used to replace the pre-trained SSD body detector in the BlazePose preheating module. Simultaneously, an offset vector is proposed to update the bounding box obtained. This method can reduce the frequency of repeated use of the head detection module. The algorithm then uses the long short-term memory neural network to detect seafarer falls. After extracting fall and related behavior data from the URFall public data set and FDD public data set to enrich the self-made data set, the experimental results show that the algorithm can achieve 100% accuracy and 98.5% specificity for the seafarer’s falling behavior, indicating that the algorithm has reasonable practicability and strong generalization ability. The detection frame rate can reach 29 fps on a CPU, which can meet the effect of real-time detection. The proposed method can be deployed on common vision sensors.

Plant Disease Detection and Classification Method Based on the Optimized Lightweight YOLOv5 Model

Traditional plant disease diagnosis methods are mostly based on expert diagnosis, which easily leads to the backwardness of crop disease control and field management. In this paper, to improve the speed and accuracy of disease classification, a plant disease detection and classification method based on the optimized lightweight YOLOv5 model is proposed. We propose an IASM mechanism to improve the accuracy and efficiency of the model, to achieve model weight reduction through Ghostnet and WBF structure, and to combine BiFPN and fast normalization fusion for weighted feature fusion to speed up the learning efficiency of each feature layer. To verify the effect of the optimized model, we conducted a performance comparison test and ablation test between the optimized model and other mainstream models. The results show that the operation time and accuracy of the optimized model are 11.8% and 3.98% higher than the original model, respectively, while F1 score reaches 92.65%, which highlight statistical metrics better than the current mainstream models. Moreover, the classification accuracy rate on the self-made dataset reaches 92.57%, indicating the effectiveness of the plant disease classification model proposed in this paper, and the transfer learning ability of the model can be used to expand the application scope in the future.

A Novel AI-Based Visual Stimuli Generation Approach for Environment Concept Design

The nature of environment concept design is a visual-based issue, where designers need to find loads of visual stimuli to create the high-quality concept. This paper aims to introduce a novel AI-based method to automatically generate images as design inspiration for environment concept design. Through interviewing eight professionals, we discovered that acquiring the design stimulus with inspired “composition” in a short time is the prioritized need of designers. This paper takes spectacular ambience for example. Through testing six classic GAN model variants trained by a self-made data set, we selected qualified four models to generate black and white thumbnails as stimuli with spectacular ambience. Moreover, we conducted a qualitative study of the outputs of the four models in a manner that invite eight designers to discuss. Finally, we summarized five key factors that influence designers' satisfaction with generated visual stimuli and discuss future directions that are worth studying.

A Novel Effective Vehicle Detection Method Based on Swin Transformer in Hazy Scenes

Under bad weather, the ability of intelligent vehicles to perceive the environment accurately is an important research content in many practical applications such as smart cities and unmanned driving. In order to improve vehicle environment perception technology in real hazy scenes, we propose an effective detection algorithm based on Swin Transformer for hazy vehicle detection. This algorithm includes two aspects. First of all, for the aspect of the difficulty in extracting haze features with poor visibility, a dehazing network is designed to obtain high-quality haze-free output through encoding and decoding methods using Swin Transformer blocks. In addition, for the aspect of the difficulty of vehicle detection in hazy images, a new end-to-end vehicle detection model in hazy days is constructed by fusing the dehazing module and the Swin Transformer detection module. In the training stage, the self-made dataset Haze-Car is used, and the haze detection model parameters are initialized by using the dehazing model and Swin-T through transfer learning. Finally, the final haze detection model is obtained by fine tuning. Through the joint learning of dehazing and object detection and comparative experiments on the self-made real hazy image dataset, it can be seen that the detection performance of the model in real-world scenes is improved by 12.5%.

A Review of Image Processing Techniques for Deepfakes.

Deep learning is used to address a wide range of challenging issues including large data analysis, image processing, object detection, and autonomous control. In the same way, deep learning techniques are also used to develop software and techniques that pose a danger to privacy, democracy, and national security. Fake content in the form of images and videos using digital manipulation with artificial intelligence (AI) approaches has become widespread during the past few years. Deepfakes, in the form of audio, images, and videos, have become a major concern during the past few years. Complemented by artificial intelligence, deepfakes swap the face of one person with the other and generate hyper-realistic videos. Accompanying the speed of social media, deepfakes can immediately reach millions of people and can be very dangerous to make fake news, hoaxes, and fraud. Besides the well-known movie stars, politicians have been victims of deepfakes in the past, especially US presidents Barak Obama and Donald Trump, however, the public at large can be the target of deepfakes. To overcome the challenge of deepfake identification and mitigate its impact, large efforts have been carried out to devise novel methods to detect face manipulation. This study also discusses how to counter the threats from deepfake technology and alleviate its impact. The outcomes recommend that despite a serious threat to society, business, and political institutions, they can be combated through appropriate policies, regulation, individual actions, training, and education. In addition, the evolution of technology is desired for deepfake identification, content authentication, and deepfake prevention. Different studies have performed deepfake detection using machine learning and deep learning techniques such as support vector machine, random forest, multilayer perceptron, k-nearest neighbors, convolutional neural networks with and without long short-term memory, and other similar models. This study aims to highlight the recent research in deepfake images and video detection, such as deepfake creation, various detection algorithms on self-made datasets, and existing benchmark datasets.

Real-Time Object Detection for the Running Train Based on the Improved YOLO V4 Neural Network

Matching the detection speed and accuracy in practical applications is considered to improve the speed of video object detection in front of trains. Lightweight convolutional neural network MobileNet and clustering ideas are combined to improve the object detection algorithm, and the MYOLO-lite model object detection algorithm is designed. The self-made object dataset of a forward-moving train is combined with the K-means clustering idea. The transcendental frame is redesigned to enhance scale adaptability. The nonmaximum suppression and loss function improvement methods of the MYOLO-lite network model are proposed for occluded tracks in front of the train, obstruction of trains, low detection accuracy of large object coincidence, and uneven distribution of positive and negative samples. The mean average precision value of the experimentally designed MYOLO-lite model object detection algorithm reaches 95.74%, with 42.04 frames per second. Detecting a picture only takes 0.024 s.

1D-DRSETL: a novel unsupervised transfer learning method for cross-condition fault diagnosis of rolling bearing

Transfer learning can meet the challenge of cross-condition fault diagnosis. However, the diagnostic effectiveness of transfer learning in actual applications is unsatisfactory, mainly due to the great unbalance in labeling between testing and training samples. A one-dimensional dual residual squeeze-and-excitation transfer learning network (1D-DRSETL) is proposed for an unsupervised accurate intelligent diagnosis under cross-condition in this paper for unlabeled small sample. First, a special block is designed to obtain transferable features by adaptively focusing on fault-sensitive information. Second, the joint maximum mean discrepancy is utilized to deal with the feature matching problem under cross-conditions. Then, speed up model training with AdaBelief optimizer. Finally, cross-conditions transfer diagnosis experiments are designed to demonstrate the superiority of the method based on a self-made dataset and the publicly available rolling bearings dataset. The experimental results show that the proposed method can achieve higher fault diagnosis accuracy and better robustness under cross-conditions than the contrasting methods.

LETR: An End-to-End Detector of Reconstruction Area in Blades Adaptive Machining with Transformer

In the leading/trailing edge’s adaptive machining of the near-net-shaped blade, a small portion of the theoretical part, called the reconstruction area, is retained for securing aerodynamic performance by manual work. The next work is to recognize the reconstruction area of the reconstructed leading/trailing edge’s image. To accelerate this process, an anchor-free neural network model based on Transformer was proposed, named Leading/trailing Edge Transformer (LETR). LETR extracts image features from an aspect of mixed frequency and channel domain. We also integrated LETR with the newest meta-Acon activation function. We tested our model on the self-made dataset LDEG2021 on a single GPU and got an mAP of 91.9%, which surpassed our baseline model, Deformable DETR, by 1.1%. Furthermore, we modified LETR’s convolution layer and named the new model after Ghost Leading/trailing Edge Transformer (GLETR) as a lightweight model for real-time detection. It is proved that GLETR has fewer weight parameters and converges faster than LETR with an acceptable decrease in mAP (0.1%) by test results. The proposed models provide the basis for subsequent parameter extraction work in the reconstruction area.

LDS-YOLO: A lightweight small object detection method for dead trees from shelter forest

The detection and location of dead trees are extremely important for the management and estimating naturalness of the forests, and timely replanting of dead trees can effectively resist natural disasters and maintain the stability of the ecosystem. Dead trees have the characteristics of small targets and inconspicuous detail information, which leads to the problem of difficult identification. In this paper, we propose a novel lightweight architecture for small objection detection based on the YOLO framework, named LDS-YOLO. Specifically, a novel feature extraction module is proposed, it reuses the features from previous layers for the purpose of dense connectivity and reduced dependence on the dataset. Then, for Spatial pyramid pooling (SPP) with the introduction of SoftPool method for retaining detailed information about the object to ensure that small targets are not missed. In the meantime, a depth-wise separable convolution with a small number of parameters is used instead of the traditional convolution to reduce the number of model parameters. We evaluate the proposed method on our self-made dataset based UAV captured images. The experimental results demonstrate that the LDS-YOLO architecture performs well in comparison with the state-of-the-art models, with AP of 89.11% and parameter size of 7.6 MB, and can be used for rapid detection of dead trees in shelter forests, which provides a scientific theoretical basis for forestry management of Three North shelter Forest.

An Object Detection Model for Paint Surface Detection Based on Improved YOLOv3

To solve the problem of poor performance of the target detection algorithm and false detection in the detection of paint surface defects of office chairs five-star feet, we propose a defect detection method based on the improved YOLOv3 algorithm. Firstly, a new feature fusion structure is designed to reduce the missed detection rate of small targets. Then we used the CIOU loss function to improve the positioning accuracy. At the same time, a parallel version of the k-means++ initialization algorithm (K-means||) is used to optimize and determine the parameters of the a priori anchor so as to improve the matching degree between the a priori anchor and the feature layer. We constructed a dataset of paint surface defects on the five-star feet of office chairs and performed optimization training, and used multiple algorithms and different datasets to conduct comparative experiments to validate the algorithm. The experimental results show that the improved YOLOv3 algorithm is effective in that the average precision on the self-made dataset reaches 88.3%, which is 5.8% higher than the original algorithm. At the same time, it has also been verified based on the Aliyun Tianchi competition aluminum dataset, and the average precision has reached 89.2%. This method realizes the real-time detection of the paint surface defects of the five-star feet of the office chair very well.

Gesture recognition based on modified Yolov5s

With the development of artificial intelligence technology, human–computer interaction technology through gestures, images and voices has gradually become a hot topic for discussion. A modified Yolov5s gesture recognition method is proposed in the field of human–computer cooperation by optimizing the network structure of Yolov5s backbone, CNN is replaced by Ghostbottleneck module to increase the target occlusion recognition rate. Secondly, tensor stitching is added to the output of Ghostbottleneck module for up sampling to strengthen the reuse of image features. Finally, the detection ability of the improved model in the face of complex environment is verified on the self-made data set. Experimental results show that, the mAP@0.5 (mean average precision) of the modified Yolov5s is 94.49%, the AP (average precision) is 94.2%. By comparing the Yolov5s algorithm, Yolov4 algorithm,Yolov3 algorithm and SSD algorithm, the detection accuracy of the modified method has been significantly improved, which can fully meet the application requirements of real-time detection of gesture-controlled robots.

Sign Language Recognition

This paper focuses on experimenting with different segmentation approaches and unsupervised learning algorithms to create an accurate sign language recognition model. To more easily approach the problem and obtain reasonable results, we experimented with just up to 10 different classes/letters in our self-made dataset instead of all 26 possible letters. We collected 12000 RGB images and their corresponding depth data using a Microsoft Kinect. Up to half of the data was fed into the autoencoder to extract features while the other half was used for testing. We achieved a classification accuracy of 98% on a randomly selected set of test data using our trained model. In addition to the work we did on static images, we also created a live demo version of the project which can be run at a little less than 2 seconds per frame to classify signed hand gestures from any person.

Cuffless Blood Pressure Estimation Based on Monte Carlo Simulation Using Photoplethysmography Signals.

Blood pressure measurements are one of the most routinely performed medical tests globally. Blood pressure is an important metric since it provides information that can be used to diagnose several vascular diseases. Conventional blood pressure measurement systems use cuff-based devices to measure the blood pressure, which may be uncomfortable and sometimes burdensome to the subjects. Therefore, in this study, we propose a cuffless blood pressure estimation model based on Monte Carlo simulation (MCS). We propose a heterogeneous finger model for the MCS at wavelengths of 905 nm and 940 nm. After recording the photon intensities from the MCS over a certain range of blood pressure values, the actual photoplethysmography (PPG) signals were used to estimate blood pressure. We used both publicly available and self-made datasets to evaluate the performance of the proposed model. In case of the publicly available dataset for transmission-type MCS, the mean absolute errors are 3.32 ± 6.03 mmHg for systolic blood pressure (SBP), 2.02 ± 2.64 mmHg for diastolic blood pressure (DBP), and 1.76 ± 2.8 mmHg for mean arterial pressure (MAP). The self-made dataset is used for both transmission- and reflection-type MCSs; its mean absolute errors are 2.54 ± 4.24 mmHg for SBP, 1.49 ± 2.82 mmHg for DBP, and 1.51 ± 2.41 mmHg for MAP in the transmission-type case as well as 3.35 ± 5.06 mmHg for SBP, 2.07 ± 2.83 mmHg for DBP, and 2.12 ± 2.83 mmHg for MAP in the reflection-type case. The estimated results of the SBP and DBP satisfy the requirements of the Association for the Advancement of Medical Instrumentation (AAMI) standards and are within Grade A according to the British Hypertension Society (BHS) standards. These results show that the proposed model is efficient for estimating blood pressures using fingertip PPG signals.

Power Grid Field Violation Recognition Algorithm Based on Enhanced YOLOv5

Aiming at the problem that Yolov5 is difficult to accurately detect backlighting and small samples in national grid power transmission and distribution operations, studies the illegal behavior technology of “two wear and one wear” in power grid operation, and proposes a violation recognition algorithm based on the enhanced YOLOv5. First, add a new detection layer and use the BiFPN (bi-directional feature pyramid network) layer for feature fusion, so that feature layers of different scales can better learn the weight distribution and enhance the fusion ability. Secondly, add CBAM (Convolutional Block Attention Module) module before the output detection layer feature map to make full use of channel and spatial information to achieve better model accuracy and recall. The experimental results on the six self-made data sets show that the mAP of this method is 91.0%, which is 5.6% higher than the original algorithm, indicating that the model has stronger predictive ability and robustness for the identification of personnel and safety tools in transmission and distribution scenarios.

Automatic generation of architecture facade for historical urban renovation using generative adversarial network

In architecture, urban culture inheritance plays a nontrivial role in the urban renovation of a city. The existing design workflow for architects is laborious from the research phase to the final plans, and yet the results are subjective, especially when facing a large-scale urban renovation. Recent works show that generative adversarial networks (GANs) have surprising potential for creating un-presented images following certain rules under various domains, including architecture. However, research on adopting GANs to optimize the conventional design process of historic urban areas renovation is still missing. This study contributes as a decision support tool for urban renovation based on GANs to abstract historic architecture styles and automatically generate stylized facades. One self-made dataset of facade and label information from Harbin Central Street was created, and one data augmentation process was proposed and examined. The generated designs were evaluated quantitatively and qualitatively, and demonstrated high accuracy, reality, and diversity levels. Two in situ and off-site applications proved the feasibility and adaptability of the proposed workflow. Through all evaluations and application cases, it is justified that the proposed GAN-based design strategy can greatly benefit the conventional historic urban area renovation design process.

CLFormer: A Lightweight Transformer Based on Convolutional Embedding and Linear Self-Attention With Strong Robustness for Bearing Fault Diagnosis Under Limited Sample Conditions

As a rising star in the field of deep learning, the Transformers have achieved remarkable achievements in numerous tasks. Nonetheless, due to the safety considerations, complex environment, and limitation of deployment cost in actual industrial production, the algorithms used for fault diagnosis often face the three challenges of limited samples, noise interference, and lightweight, which is an impediment in the fault diagnosis practice of transformer with high requirements for number of samples and parameters. For this reason, this article proposes a lightweight transformer based on convolutional embedding and linear self-attention (LSA), called CLFormer. By modifying the embedding module and the form of self-attention, the aim of lightweight is realized (MFLOPs: 0.12; Params: 4.88 K) under the condition of boosting high-accuracy of transformer. The effectiveness was demonstrated on Self-Made dataset with four comparative models, especially when each type of training sample is <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$6\times 4$ </tex-math></inline-formula> , the CLFormer achieves the highest average accuracy of 83.58% when the signal-to-noise ratio (SNR) is from −8 to 8 dB for three types of noise. As the first attempt to use transformer for fault diagnosis of rotating machinery, this work provides a feasible strategy for the research topic of fault diagnosis with the goal of practical deployment.

An automatic sorting system for electronic components detached from waste printed circuit boards.

Recycling e-waste makes for eliminating the pollution to environment and recovering critical materials as one of resource. Printed circuit boards (PCBs) serve as the important part in all e-waste, containing valuable but hazardous elements to be recycled when they reach the end of life. For the recycling of waste PCB, the electronic components (ECs) are liberated from base board and to be treated separately for element recovery. Due to the diverse element composition, ECs deserve to be further classified and sorted to improve the efficiency of recycling, achieving the concept of accurate recovery. Currently, the recycling industry only roughly screen the ECs manually by labors, which increases the risk of health for exposure to the hazardous environment. Automatic solutions are necessary for replacing labors to classify and sort the waste ECs, thus safeguarding them against the hazards of factory environment. In this work, the YOLO-V3, an emerging image detection algorithm, is utilized to train the self-made dataset and classify the ECs into specific categories. To avoid surface damage that weakens the accuracy of object detection, the technology process of detaching the ECs is improved by building a nitrogen atmosphere during the desoldering process, which delivers great protection effects on ECs. Results of YOLO-V3 detection model present satisfactory classification capability for all the classes of ECs and a smart on-line sorting system is proposed to automatically separate the ECs detached from WPCB.

Vehicle Navigation System based on Pollution Metric Analysis with Q-Learning Algorithm

The navigation systems available in the present scenario takes into account the path distance for their estimations. In some advanced navigation systems, the road traffic analysis is also considered in the algorithm for their predictions. The proposed work estimates a navigation path with respect to the present pollution level on the roadways. The work suggests an alternate path to avoid additional vehicles to enter the same road which is already impacted by air pollution. A Q-learning (Quality learning) prediction algorithm is trained in the proposed work with a self-made dataset for the estimations. The experimental work presented in the paper explores the accuracy and computational speed of the developed algorithm in comparison to the traditional algorithms.