基于特征金字塔的多尺度特征融合网络

特征金字塔网络（FPN）是CNN网络对图像信息进行表达输出的一种有效方法,在目标检测网络中广泛应用．然而,FPN没有充分地将浅层的细节信息传递到深层的语义特征,存在特征融合不足的缺陷,因而只能依靠深层语义信息来进行预测,从而忽略了网络低层细节信息,对各种视觉学习的效果造成了一定的影响．针对FPN存在的以上问题,本文提出基于特征金字塔的多尺度特征融合网络模型,在FPN主干网络的基础上,设计了混合特征金字塔和金字塔融合模块,并结合注意力机制,对特征金字塔进行了多尺度的深度融合．本文在PASCAL VOC2012和MS COCO2014数据集上,以Faster R-CNN作为基础检测器进行实验,验证了MFPN对特征融合的有效性．     

Severity detection and infection level identification of tuberculosis using deep learning

Tuberculosis (TB) is a highly infectious disease and is one of the major health problems all over the world. The accurate detection of TB is a major challenge faced by most of the existing methods. This work addresses these issues and developed an effective mechanism for detecting TB using deep learning. Here, the color space transformation is applied for transforming the red green and blue image to LUV space, where L stands for luminance, U and V represent chromaticity values of color images. Then, adaptive thresholding is carried out for image segmentation and various features, like coverage, density, color histogram, area, length, and texture features, are extracted to enable effective classification. After the feature extraction, the size of the features is reduced using principal component analysis. The extracted features are subjected to fractional crow search-based deep convolutional neural network (FC-SVNN) for the classification. Then, the image level features, like bacilli count, bacilli area, scattering coefficients and skeleton features are considered to perform severity detection using proposed adaptive fractional crow (AFC)-deep CNN. Finally, the inflection level is determined using entropy, density and detection percentage. The proposed AFC-Deep CNN algorithm is designed by modifying FC algorithm using self-adaptive concept. The proposed AFC-Deep CNN shows better performance with maximum accuracy value as 0.935.     

基于深度学习的烟包识别与分类

目的 提取烟包图像数据训练深度学习目标检测模型,提升烟包流水线拣包效率和准确性。方法 基于深度学习建立一种烟包识别分类模型,对原始YOLOv3模型进行改进,在原网络中加入设计的多空间金字塔池化结构(M–SPP),将64×64尺度的特征图下采样与32×32尺度的特征图进行拼接,并去除16×16尺度的预测特征层,提高模型的检测准确率和速度,并采用K–means++算法对先验框参数进行优化。结果 实验表明该目标检测模型平均准确率达到99.68%,检测速度达到70.82帧/s。结论 基于深度学习建立的图像识别分类模型准确率高且检测速度快,有效满足烟包流水线自动化实时检测。     

IoT & AI Enabled Three-Phase Secure and Non-Invasive COVID 19 Diagnosis System

Corona is a viral disease that has taken the form of an epidemic and is causing havoc worldwide after its first appearance in the Wuhan state of China in December 2019. Due to the similarity in initial symptoms with viral fever, it is challenging to identify this virus initially. Non-detection of this virus at the early stage results in the death of the patient. Developing and densely populated countries face a scarcity of resources like hospitals, ventilators, oxygen, and healthcare workers. Technologies like the Internet of Things (IoT) and artificial intelligence can play a vital role in diagnosing the COVID-19 virus at an early stage. To minimize the spread of the pandemic, IoT-enabled devices can be used to collect patient’s data remotely in a secure manner. Collected data can be analyzed through a deep learning model to detect the presence of the COVID-19 virus. In this work, the authors have proposed a three-phase model to diagnose covid-19 by incorporating a chatbot, IoT, and deep learning technology. In phase one, an artificially assisted chatbot can guide an individual by asking about some common symptoms. In case of detection of even a single sign, the second phase of diagnosis can be considered, consisting of using a thermal scanner and pulse oximeter. In case of high temperature and low oxygen saturation levels, the third phase of diagnosis will be recommended, where chest radiography images can be analyzed through an AI-based model to diagnose the presence of the COVID-19 virus in the human body. The proposed model reduces human intervention through chatbot-based initial screening, sensor-based IoT devices, and deep learning-based X-ray analysis. It also helps in reducing the mortality rate by detecting the presence of the COVID-19 virus at an early stage.     

Multimodal brain tumor detection and classification using deep saliency map and improved dragonfly optimization algorithm

In the last decade, there has been a significant increase in medical cases involving brain tumors. Brain tumor is the tenth most common type of tumor, affecting millions of people. However, if it is detected early, the cure rate can increase. Computer vision researchers are working to develop sophisticated techniques for detecting and classifying brain tumors. MRI scans are primarily used for tumor analysis. We proposed an automated system for brain tumor detection and classification using a saliency map and deep learning feature optimization in this paper. The proposed framework was implemented in stages. In the initial phase of the proposed framework, a fusion-based contrast enhancement technique is proposed. In the following phase, a tumor segmentation technique based on saliency maps is proposed, which is then mapped on original images based on active contour. Following that, a pre-trained CNN model named EfficientNetB0 is fine-tuned and trained in two ways: on enhanced images and on tumor localization images. Deep transfer learning is used to train both models, and features are extracted from the average pooling layer. The deep learning features are then fused using an improved fusion approach known as Entropy Serial Fusion. The best features are chosen in the final step using an improved dragonfly optimization algorithm. Finally, the best features are classified using an extreme learning machine (ELM). The experimental process is conducted on three publically available datasets and achieved an improved accuracy of 95.14, 94.89, and 95.94%, respectively. The comparison with several neural nets shows the improvement of proposed framework.     

A Hybrid Attention-Based Residual Unet for Semantic Segmentation of Brain Tumor

Segmenting brain tumors in Magnetic Resonance Imaging (MRI) volumes is challenging due to their diffuse and irregular shapes. Recently, 2D and 3D deep neural networks have become famous for medical image segmentation because of the availability of labelled datasets. However, 3D networks can be computationally expensive and require significant training resources. This research proposes a 3D deep learning model for brain tumor segmentation that uses lightweight feature extraction modules to improve performance without compromising contextual information or accuracy. The proposed model, called Hybrid Attention-Based Residual Unet (HA-RUnet), is based on the Unet architecture and utilizes residual blocks to extract low- and high-level features from MRI volumes. Attention and Squeeze-Excitation (SE) modules are also integrated at different levels to learn attention-aware features adaptively within local and global receptive fields. The proposed model was trained on the BraTS-2020 dataset and achieved a dice score of 0.867, 0.813, and 0.787, as well as a sensitivity of 0.93, 0.88, and 0.83 for Whole Tumor, Tumor Core, and Enhancing Tumor, on test dataset respectively. Experimental results show that the proposed HA-RUnet model outperforms the ResUnet and AResUnet base models while having a smaller number of parameters than other state-of-the-art models. Overall, the proposed HA-RUnet model can improve brain tumor segmentation accuracy and facilitate appropriate diagnosis and treatment planning for medical practitioners.     

基于改进YOLOv5的轻量级芯片封装缺陷检测方法

目的 针对芯片封装缺陷检测过程中检测精度低与模型难部署的问题,提出YOLOv5-SPM检测网络,旨在提高检测精度并实现模型轻量化。方法 首先,通过在特征提取模块后增加通道注意力机制,提高缺陷通道的关注度,减少冗余特征的干扰,进而提升目标的检测精度。其次,在主干网络与颈部网络连接处使用快速特征金字塔结构,更好地融合了自建芯片数据集的多尺度特征信息。最后,将主干网络的特征提取模块更换为MobileNetV3,将常规卷积更换为深度卷积和点卷积,有效降低了模型尺寸和计算量。结果 经过改进后的新网络YOLOv5s-SPM在模型参数下降29.5%的情况下,平均精度较原网络提高了0.6%,准确率提高了3.2%。结论 新网络相较于传统网络在芯片缺陷检测任务中实现了模型精度与速度的统一提高,同时由于模型参数减小了29.5%,更适合部署在资源有限的工业嵌入式设备上。     

基于视觉感知的机器人工件识别方法研究

目的 解决定制化木门尺寸规格不统一、表面纹理多样而导致的堆垛分类困难、搬运效率低下等问题。方法 提出采用深度学习方法进行定制式木门工件检测,以YOLOV3网络为基本框架开展机器人工件识别方法研究。首先,通过图像数据增强和预处理,扩充定制式木门数据;然后,进行YOLO V3损失函数改进,并根据木门特征进行定制式木门数据集锚框尺度的重新聚类;最后,应用空间金字塔池化层进行YOLO V3中特征金字塔网络改进,并通过随机选取的测试集验证本文方法的有效性。结果 测试数据集的平均检测准确率均值达到98.05%,检测每张图片的时间为137 ms。结论 研究表明,本文方法能够满足木门生产线对准确率和实时性的要求,可大大提高定制化木门转线及堆垛效率。     

Developing a Recognition System for Classifying COVID-19 Using a Convolutional Neural Network Algorithm

The COVID-19 pandemic poses an additional serious public health threat due to little or no pre-existing human immunity, and developing a system to identify COVID-19 in its early stages will save millions of lives. This study applied support vector machine (SVM), k-nearest neighbor (K-NN) and deep learning convolutional neural network (CNN) algorithms to classify and detect COVID-19 using chest X-ray radiographs. To test the proposed system, chest X-ray radiographs and CT images were collected from different standard databases, which contained 95 normal images, 140 COVID-19 images and 10 SARS images. Two scenarios were considered to develop a system for predicting COVID-19. In the first scenario, the Gaussian filter was applied to remove noise from the chest X-ray radiograph images, and then the adaptive region growing technique was used to segment the region of interest from the chest X-ray radiographs. After segmentation, a hybrid feature extraction composed of 2D-DWT and gray level co-occurrence matrix was utilized to extract the features significant for detecting COVID-19. These features were processed using SVM and K-NN. In the second scenario, a CNN transfer model (ResNet 50) was used to detect COVID-19. The system was examined and evaluated through multiclass statistical analysis, and the empirical results of the analysis found significant values of 97.14%, 99.34%, 99.26%, 99.26% and 99.40% for accuracy, specificity, sensitivity, recall and AUC, respectively. Thus, the CNN model showed significant success; it achieved optimal accuracy, effectiveness and robustness for detecting COVID-19.     

基于多层次视觉语义特征融合的图像检索算法

目的为了解决低层特征与中层语义属性间出现的语义鸿沟,以及在将低层特征转化为语义属性的过程中易丢失信息,从而会降低检索精度等问题,设计一种多层次视觉语义特征融合的图像检索算法。方法首先分别提取图像的3种中层特征(深度卷积神经网络(DCNN)特征、Fisher向量、稀疏编码空间金字塔匹配特征(SCSPM));其次,为了对3种特征进行有效融合,定义一种基于图的半监督学习模型,将提取的3个中层特征进行融合,形成一个多层次视觉语义特征,有效结合3种不同中层特征的互补信息,提高图像特征描述,从而降低检索算法中的语义鸿沟;最后,引入具有视觉特性与语义统一的距离函数,根据提取的多层次视觉语义特征来计算查询图像和训练图像的相似度量,完成图像检索任务。结果实验结果表明,与当前检索方法对比,文中算法具有更高的检索精度与效率。结论所提算法具有良好的检索准确度,在医疗、包装商标等领域具有一定的参考价值。     

A Lightweight Convolutional Neural Network with Representation Self-challenge for Fingerprint Liveness Detection

Fingerprint identification systems have been widely deployed in many occasions of our daily life. However, together with many advantages, they are still vulnerable to the presentation attack (PA) by some counterfeit fingerprints. To address challenges from PA, fingerprint liveness detection (FLD) technology has been proposed and gradually attracted people's attention. The vast majority of the FLD methods directly employ convolutional neural network (CNN), and rarely pay attention to the problem of over-parameterization and over-fitting of models, resulting in large calculation force of model deployment and poor model generalization. Aiming at filling this gap, this paper designs a lightweight multi-scale convolutional neural network method, and further proposes a novel hybrid spatial pyramid pooling block to extract abundant features, so that the number of model parameters is greatly reduced, and support multi-scale true/fake fingerprint detection. Next, the representation self-challenge (RSC) method is used to train the model, and the attention mechanism is also adopted for optimization during execution, which alleviates the problem of model over-fitting and enhances generalization of detection model. Finally, experimental results on two publicly benchmarks: LivDet2011 and LivDet2013 sets, show that our method achieves outstanding detection results for blind materials and cross-sensor. The size of the model parameters is only 548 KB, and the average detection error of cross-sensors and cross-materials are 15.22 and 1 respectively, reaching the highest level currently available.     

An IoT-Cloud Based Intelligent Computer-Aided Diagnosis of Diabetic Retinopathy Stage Classification Using Deep Learning Approach

Diabetic retinopathy (DR) is a disease with an increasing prevalence and the major reason for blindness among working-age population. The possibility of severe vision loss can be extensively reduced by timely diagnosis and treatment. An automated screening for DR has been identified as an effective method for early DR detection, which can decrease the workload associated to manual grading as well as save diagnosis costs and time. Several studies have been carried out to develop automated detection and classification models for DR. This paper presents a new IoT and cloud-based deep learning for healthcare diagnosis of Diabetic Retinopathy (DR). The proposed model incorporates different processes namely data collection, preprocessing, segmentation, feature extraction and classification. At first, the IoT-based data collection process takes place where the patient wears a head mounted camera to capture the retinal fundus image and send to cloud server. Then, the contrast level of the input DR image gets increased in the preprocessing stage using Contrast Limited Adaptive Histogram Equalization (CLAHE) model. Next, the preprocessed image is segmented using Adaptive Spatial Kernel distance measure-based Fuzzy C-Means clustering (ASKFCM) model. Afterwards, deep Convolution Neural Network (CNN) based Inception v4 model is applied as a feature extractor and the resulting feature vectors undergo classification in line with the Gaussian Naive Bayes (GNB) model. The proposed model was tested using a benchmark DR MESSIDOR image dataset and the obtained results showcased superior performance of the proposed model over other such models compared in the study.     

A Model for Helmet-Wearing Detection of Non-Motor Drivers Based on YOLOv5s

Detecting non-motor drivers’ helmets has significant implications for traffic control. Currently, most helmet detection methods are susceptible to the complex background and need more accuracy and better robustness of small object detection, which are unsuitable for practical application scenarios. Therefore, this paper proposes a new helmet-wearing detection algorithm based on the You Only Look Once version 5 (YOLOv5). First, the Dilated convolution In Coordinate Attention (DICA) layer is added to the backbone network. DICA combines the coordinated attention mechanism with atrous convolution to replace the original convolution layer, which can increase the perceptual field of the network to get more contextual information. Also, it can reduce the network’s learning of unnecessary features in the background and get attention to small objects. Second, the Rebuild Bidirectional Feature Pyramid Network (Re-BiFPN) is used as a feature extraction network. Re-BiFPN uses cross-scale feature fusion to combine the semantic information features at the high level with the spatial information features at the bottom level, which facilitates the model to learn object features at different scales. Verified on the proposed “Helmet Wearing dataset for Non-motor Drivers (HWND),” the results show that the proposed model is superior to the current detection algorithms, with the mean average precision (mAP) of 94.3% under complex background.     

基于改进BiSeNet的室内场景语义分割方法

室内场景的语义分割一直是深度学习语义分割领域的一个重要方向。室内语义分割主要存在的问题有语义类别多、很多物体类会有相互遮挡、某些类之间相似性较高等。针对这些问题,提出了一种用于室内场景语义分割的方法。该方法在BiSeNet(bilateral segmentation network)的网络结构基础上,引入了一个空洞金字塔池化层和多尺度特征融合模块,将上下文路径中的浅层细节特征与通过空洞金字塔池化得到的深层抽象特征进行融合,得到增强的内容特征,提高模型对室内场景语义分割的表现。该方法在ADE20K中关于室内场景的数据集上的MIoU表现,比SegNet高出23.5%,比改进前高出3.5%。     

序列图像中运动目标检测

提出动态背景下序列图像中的运动目标检测算法。利用像素邻域的各向同性对图像进行归一化,消除亮度变化等因素的影响;利用光流信息并结合小波变换由粗及精计算速度场来配准图像;用当前帧作参考图像,通过时域积分校正背景图像。当前帧与校正后背景图像作差得到差分图像。假设该差分图像中噪声分布为高斯分布,由高斯分布的3σ特性滤除差分图像中的噪声,则粗定位出目标;最后以聚类方法确定运动目标区域。分别对200帧可见光和200帧红外图像序列进行实验,检测率分别为95%和94%。     

Road Damage Detection and Classification Using Mask R-CNN with DenseNet Backbone

Automatic road damage detection using image processing is an important aspect of road maintenance. It is also a challenging problem due to the inhomogeneity of road damage and complicated background in the road images. In recent years, deep convolutional neural network based methods have been used to address the challenges of road damage detection and classification. In this paper, we propose a new approach to address those challenges. This approach uses densely connected convolution networks as the backbone of the Mask R-CNN to effectively extract image feature, a feature pyramid network for combining multiple scales features, a region proposal network to generate the road damage region, and a fully convolutional neural network to classify the road damage region and refine the region bounding box. This method can not only detect and classify the road damage, but also create a mask of the road damage. Experimental results show that the proposed approach can achieve better results compared with other existing methods.     

基于卷积神经网络模型的遥感图像分类

研究了遥感图像的分类,针对遥感图像的支持向量机(SVM)等浅层结构分类模型特征提取困难、分类精度不理想等问题,设计了一种卷积神经网络(CNN)模型,该模型包含输入层、卷积层、全连接层以及输出层,采用Soft Max分类器进行分类。选取2010年6月6日Landsat TM5富锦市遥感图像为数据源进行了分类实验,实验表明该模型采用多层卷积池化层能够有效地提取非线性、不变的地物特征,有利于图像分类和目标检测。针对所选取的影像,该模型分类精度达到94.57%,比支持向量机分类精度提高了5%,在遥感图像分类中具有更大的优势。     

Semantic Segmentation by Using Down-Sampling and Subpixel Convolution: DSSC-UNet

Recently, semantic segmentation has been widely applied to image processing, scene understanding, and many others. Especially, in deep learning-based semantic segmentation, the U-Net with convolutional encoder-decoder architecture is a representative model which is proposed for image segmentation in the biomedical field. It used max pooling operation for reducing the size of image and making noise robust. However, instead of reducing the complexity of the model, max pooling has the disadvantage of omitting some information about the image in reducing it. So, this paper used two diagonal elements of down-sampling operation instead of it. We think that the down-sampling feature maps have more information intrinsically than max pooling feature maps because of keeping the Nyquist theorem and extracting the latent information from them. In addition, this paper used two other diagonal elements for the skip connection. In decoding, we used Subpixel Convolution rather than transposed convolution to efficiently decode the encoded feature maps. Including all the ideas, this paper proposed the new encoder-decoder model called Down-Sampling and Subpixel Convolution U-Net (DSSC-UNet). To prove the better performance of the proposed model, this paper measured the performance of the U-Net and DSSC-UNet on the Cityscapes. As a result, DSSC-UNet achieved 89.6% Mean Intersection Over Union (Mean-IoU) and U-Net achieved 85.6% Mean-IoU, confirming that DSSC-UNet achieved better performance.     

An Image Edge Detection Algorithm Based on Multi-Feature Fusion

Edge detection is one of the core steps of image processing and computer vision. Accurate and fine image edge will make further target detection and semantic segmentation more effective. Holistically-Nested edge detection (HED) edge detection network has been proved to be a deep-learning network with better performance for edge detection. However, it is found that when the HED network is used in overlapping complex multi-edge scenarios for automatic object identification. There will be detected edge incomplete, not smooth and other problems. To solve these problems, an image edge detection algorithm based on improved HED and feature fusion is proposed. On the one hand, features are extracted using the improved HED network: the HED convolution layer is improved. The residual variable convolution block is used to replace the normal convolution enhancement model to extract features from edges of different sizes and shapes. Meanwhile, the empty convolution is used to replace the original pooling layer to expand the receptive field and retain more global information to obtain comprehensive feature information. On the other hand, edges are extracted using Otsu algorithm: Otsu-Canny algorithm is used to adaptively adjust the threshold value in the global scene to achieve the edge detection under the optimal threshold value. Finally, the edge extracted by improved HED network and Otsu-Canny algorithm is fused to obtain the final edge. Experimental results show that on the Berkeley University Data Set (BSDS500) the optimal data set size (ODS) F-measure of the proposed algorithm is 0.793; the average precision (AP) of the algorithm is 0.849; detection speed can reach more than 25 frames per second (FPS), which confirms the effectiveness of the proposed method.     

基于图像匹配的透明车灯部件检测

针对车灯总成的完整性检测,提出了基于图像金字塔和归一化互相关函数结合的分层匹配算法。利用直方图均衡化和锐化滤波增强图像对比度以及边缘细节信息,再运用归一化互相关相似度函数实现车灯透明部件的匹配,同时采用金字塔分层来提高图像匹配的速度。通过实验确定了合适的匹配窗口大小、金字塔层数、相似度阈值等,实现了车灯透明部件的安装检测。实验结果表明:该方法对车灯透明部件的成功检测率可达到约95%,检测时间约50 ms,具有较高的准确率和实时性。