引入反馈注意力的并行式多分辨率语义分割算法

目的 为了进一步提升语义分割精度,解决当前语义分割算法中特征图分辨率低下,低级信息特征随意丢弃,以及上下文重要信息不能顾及等问题,文中尝试提出一种融合反馈注意力模块的并行式多分辨率语义分割算法。方法 该算法提出一种并行式网络结构,在其中融合了高低分辨率信息,尽可能多地保留高维信息,减少低级信息要素的丢失,提升分割图像的分辨率。同时还在主干网络中嵌入了带反馈机制的感知注意力模块,从通道、空间、全局3个角度获得每个样本的权重信息,着重加强样本之间的特征重要性。在训练过程中,还使用了改进的损失函数,降低训练和优化难度。结果 经实验表明,文中的算法模型在PASCAL VOC2012、Camvid上的MIOU指标分别为77.78%、58.67%,在ADE20K上的也有42.52%,体现了出较好的分割性能。结论 文中的算法模型效果相较于之前的分割网络有一定程度的提升,算法中的部分模块嵌入别的主干网络依旧表现出较好的性能,展现了文中算法模型具备一定的有效性和泛化能力。     

基于双分支特征提取的轻量级图像分割算法

目的 为了提升彩色图像的分割精度，解决彩色图像分割中存在庞大计算成本和冗余参数的问题，本文提出一种双分支特征提取网络来解决上述问题。方法 双分支特征提取网络主要由语义信息分支和空间细节分支组成。语义信息分支通过在非对称残差模块中设置不同的空洞卷积率来获取输入图像不同尺度的上下文信息。空间细节分支是一个浅层且简单的网络，用于建立每个像素间的局部依赖关系以保留细节。在双分支之后连接一个特征聚合模块来有效地结合这2个分支的输出。结果 在没有任何预训练和后处理的情况下，在单块RTX2080Ti GPU上仅用0.91 M参数在Cityscapes数据集上以97帧/s的速度实现75.1%的分割准确性，在Camvid数据集上以107帧/s的推理速度取得了70.5%的分割效果。结论 通过大量实验证明，本文模型在分割准确性和效率之间取得了较好的平衡。     

基于DeepLabv3+的图像语义分割优化方法

目的为了实现良好的图像语义分割精度,同时尽可能降低网络的参数量,加快网络训练速度,提出基于DeepLabv3+的图像语义分割优化方法。方法编码器主干网络增加注意力机制模块,并采用更密集的特征池化模块有效聚合多尺度特征,同时使用深度可分离卷积降低网络计算复杂度。结果基于CamVid数据集的对比实验显示,优化后网络的MIoU分数达到了71.03%,在像素精度、平均像素精度等其他方面的评价指标上较原网络有小幅提升,并且网络参数量降低了12%。在Cityscapes的测试数据集上的MIoU分数为75.1%。结论实验结果表明,优化后的网络能够有效提取图像特征信息,提高语义分割精度,同时降低模型复杂度。文中网络使用城市道路场景数据集进行测试,可以为今后的无人驾驶技术的应用提供参考,具有一定的实际意义。     

基于语义分割的火车车厢位置检测研究

为了实现在煤炭定量装车站装车过程中实时检测火车车厢位置,为溜槽升降提供触发信号,设计了一种基于语义分割的火车车厢位置检测模型。以FPN （feature pyramid networks,特征金字塔网络）和ResNet101 （residual network,残差网络）为主干网络,提取并融合分辨率、语义强度不同的特征图;结合基于期望最大化（expectation maximization, EM）算法的注意力机制,构建车厢上边框语义分割模型,用于过滤特征图中的噪声,提高图像边界的语义分割精度;设计位置检测模块,计算语义分割后图像中各类别的面积及其比例和车厢上边框外接矩形高度,以获取火车车厢位置信息。结果表明,所构建的车厢上边框语义分割模型在测试集上的mIoU （mean intersection over union,均交并比）为81.21%,mPA （mean pixel accuracy,平均像素精度）为88.64%,相比未引入注意力机制的语义分割模型分别提升了3.91%和7.44%。在煤炭定量装车站现场进行的火车车厢位置检测试验结果表明,基于语义分割的火车车厢位置检测模型的检测精度满足煤炭装车过程中车厢位置检测任务的要求,这为实现煤炭定量装车系统的智能化提供了新思路。     

Fire Detection Method Based on Improved Fruit Fly Optimization-Based SVM

Aiming at the defects of the traditional fire detection methods, which are caused by false positives and false negatives in large space buildings, a fire identification detection method based on video images is proposed. The algorithm first uses the hybrid Gaussian background modeling method and the RGB color model to perform fire prejudgment on the video image, which can eliminate most non-fire interferences. Secondly, the traditional regional growth algorithm is improved and the fire image segmentation effect is effectively improved. Then, based on the segmented image, the dynamic and static features of the fire flame are further analyzed and extracted in the area of the suspected fire flame. Finally, the dynamic features of the extracted fire flame images were fused and classified by improved fruit fly optimization support vector machine, and the recognition results were obtained. The video-based fire detection method proposed in this paper greatly improves the accuracy of fire detection and is suitable for fire detection and identification in large space scenarios.     

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%。     

Cell-type based semantic segmentation of histopathological images using deep convolutional neural networks

Histopathological whole-slide image (WSI) analysis is still one of the most important ways to identify regions of cancer risk. For cancer in which early diagnosis is vital, pathologists are at the center of the decision-making process. Thanks to the widespread use of digital pathology and the development of artificial intelligence methods, automatic histopathological image analysis methods help pathologists in their decision-making process. In this process, rather than producing labels for whole-slide image patches, semantic segmentation is very useful, which facilitates the pathologists’ interpretation. In this study, automatic semantic segmentation based on cell type is proposed for the first time in the literature using novel deep convolutional networks structure (DCNN). We presents semantic information on four classes, including white areas in the whole-slide image, tissue without cells, tissue with normal cells and tissue with cancerous cells. This visual information presented to the pathologist is an easy-to-understand picture of the status of the cells and their implications for the spread of cancerous cells. A new DCNN architecture is created, inspired by the residual network and deconvolution network architecture. Our network is trained end-to-end manner with histopathological image patches for cell structures to be more discriminative. The proposed method not only produces more successful results than other state-of-art semantic segmentation algorithms with 9.2% training error and 88.89% F-score for test, but also has the most important advantage in that it has the ability to generate automatic information about the cancer and also provides information that pathologists can quickly interpret.     

基于改进Mask Scoring R-CNN的铲齿磨损检测研究

为了实现在电铲工作过程中对铲齿磨损进行实时检测，防止因铲齿磨损而影响电铲开采效率，提出了一种基于改进Mask Scoring R-CNN（region convolutional neural network，区域卷积神经网络）的铲齿实例分割模型。首先，以ResNet-101（residual network， 残差网络）和改进的FPN（feature pyramid networks，特征金字塔网络）作为主干网络，提取高、低特征层的语义信息和细节特征并融合，结合ROI Align层对局部特征层进行裁剪和归一化处理，以完成目标检测与实例分割；然后，基于获取的铲齿分割效果图以及二值化掩码图形信息，计算实例分割后图像中铲齿部分的像素面积，以判断其磨损情况。结果表明，以ResNet-101和改进FPN为主干网络的铲齿实例分割模型在测试集上的平均像素精度为90.76%，平均交并比为83.62%，相比于以ResNet-101和传统FPN为主干网络的实例分割模型分别提升了1.18%和1.21%。在电铲采掘工作现场进行8次铲齿磨损检测实验，检测到的每颗铲齿的磨损程度波动幅度均小于2%，均方差为0.7左右，说明所提出的实例分割模型对铲齿有较好的分割效果和稳定性，基本满足磨损检测要求。研究结果可为铲齿磨损状态的智能化检测提供新思路。     

Hybrid Segmentation Scheme for Skin Features Extraction Using Dermoscopy Images

Objective and quantitative assessment of skin conditions is essential for cosmeceutical studies and research on skin aging and skin regeneration. Various handcraft-based image processing methods have been proposed to evaluate skin conditions objectively, but they have unavoidable disadvantages when used to analyze skin features accurately. This study proposes a hybrid segmentation scheme consisting of Deeplab v3+ with an Inception-ResNet-v2 backbone, LightGBM, and morphological processing (MP) to overcome the shortcomings of handcraft-based approaches. First, we apply Deeplab v3+ with an Inception-ResNet-v2 backbone for pixel segmentation of skin wrinkles and cells. Then, LightGBM and MP are used to enhance the pixel segmentation quality. Finally, we determine several skin features based on the results of wrinkle and cell segmentation. Our proposed segmentation scheme achieved a mean accuracy of 0.854, mean of intersection over union of 0.749, and mean boundary F1 score of 0.852, which achieved 1.1%, 6.7%, and 14.8% improvement over the panoptic-based semantic segmentation method, respectively.     

基于注意力机制的轻量级RGB-D图像语义分割网络

目的针对卷积神经网络在RGB-D(彩色-深度)图像中进行语义分割任务时模型参数量大且分割精度不高的问题,提出一种融合高效通道注意力机制的轻量级语义分割网络。方法文中网络基于RefineNet,利用深度可分离卷积(Depthwiseseparableconvolution)来轻量化网络模型,并在编码网络和解码网络中分别融合高效的通道注意力机制。首先RGB-D图像通过带有通道注意力机制的编码器网络,分别对RGB图像和深度图像进行特征提取;然后经过融合模块将2种特征进行多维度融合;最后融合特征经过轻量化的解码器网络得到分割结果,并与RefineNet等6种网络的分割结果进行对比分析。结果对提出的算法在语义分割网络常用公开数据集上进行了实验,实验结果显示文中网络模型参数为90.41 MB,且平均交并比(mIoU)比RefineNet网络提高了1.7%,达到了45.3%。结论实验结果表明,文中网络在参数量大幅减少的情况下还能提高了语义分割精度。     

Double regularization control based on level set evolution for tablet packaging image segmentation

This paper proposes a novel double regularization control(DRC) method which is used for tablet packaging image segmentation.Since the intensities of tablet packaging images are inhomogenous,it is difficult to make image segmentation.Compared to methods based on level set,the proposed DRC method has some advantages for tablet packaging image segmentation.The local regional control term and the rectangle initialization contour are first employed in this method to quickly segment uneven grayscale images and accelerate the curve evolution rate.Gaussian filter operator and the convolution calculation are then adopted to remove the effects of texture noises in image segmentation.The developed penalty energy function,as regularization term,increases the constrained conditions based on the gradient flow conditions.Since the potential function is embedded into the level set of evolution equations and the image contour evolutions are bilaterally extended,the proposed method further improves the accuracy of image contours.Experimental studies show that the DRC method greatly improves the computational efficiency and numerical accuracy,and achieves better results for image contour segmentation compared to other level set methods.     

CANFIS based glioma brain tumor classification and retrieval system for tumor diagnosis

Brain tumor classification and retrieval system plays an important role in medical field. In this paper, an efficient Glioma Brain Tumor detection and its retrieval system is proposed. The proposed methodology consists of two modules as classification and retrieval. The classification modules are designed using preprocessing, feature extraction and tumor detection techniques using Co‐Active Adaptive Neuro Fuzzy Inference System (CANFIS) classifier. The image enhancement can be achieved using Heuristic histogram equalization technique as preprocessing and further texture features as Local Ternary Pattern (LTP) features and Grey Level Co‐occurrence Matrix (GLCM) features are extracted from the enhanced image. These features are used to classify the brain image into normal and abnormal using CANFIS classifier. The tumor region in abnormal brain image is segmented using normalized graph cut segmentation algorithm. The retrieval module is used to retrieve the similar segmented tumor regions from the dataset for diagnosing the tumor region using Euclidean algorithm. The proposed Glioma Brain tumor classification methodology achieves 97.28% sensitivity, 98.16% specificity and 99.14% accuracy. The proposed retrieval system achieves 97.29% precision and 98.16% recall rate with respect to ground truth images.     

Tissue Segmentation in Nasopharyngeal CT Images Using TwoStage Learning

Tissue segmentation is a fundamental and important task in nasopharyngeal images analysis. However, it is a challenging task to accurately and quickly segment various tissues in the nasopharynx region due to the small difference in gray value between tissues in the nasopharyngeal image and the complexity of the tissue structure. In this paper, we propose a novel tissue segmentation approach based on a two-stage learning framework and U-Net. In the proposed methodology, the network consists of two segmentation modules. The first module performs rough segmentation and the second module performs accurate segmentation. Considering the training time and the limitation of computing resources, the structure of the second module is simpler and the number of network layers is less. In addition, our segmentation module is based on U-Net and incorporates a skip structure, which can make full use of the original features of the data and avoid feature loss. We evaluated our proposed method on the nasopharyngeal dataset provided by West China Hospital of Sichuan University. The experimental results show that the proposed method is superior to many standard segmentation structures and the recently proposed nasopharyngeal tissue segmentation method, and can be easily generalized across different tissue types in various organs.     

基于差异积累的视频运动对象自动分割

针对视频运动对象的自动分割,本文给出了一种基于差异积累的自动分割算法。与传统的基于运动信息变化检测方法不同,该算法通过累积的帧差信息构建出可靠的背景,与当前帧比较进而提取出视频运动对象。本文提出了一种增强的基于Otsu法的自适应阈值化方法,能更准确地对背景差图像进行阈值化分割,克服了传统Otsu法阈值化容易失效的问题。改进的基于区域生长的定位方法更能避免传统方法的误定位及重定位的问题。实验结果表明,本文算法具有较好的实时性、自适应性以及鲁棒性,可以较为可靠地建立背景模型并进行实时更新,适用于刚体或非刚体存在平缓的光照变化以及摄像头微抖动的视频运动对象的自动分割。     

采用递归门限分析的红外目标分割

提出了一种有效的基于递归门限分析的红外目标分割方法。针对传统方法在目标的相对面积较小时背景信息容易误分的问题,将传统分割方法和递归处理结合起来,用于分割红外目标。在分割时,将每次分割得到的背景部分(即暗部分)淘汰掉,而保留分割得到的目标部分(即亮部分)。对得到的目标部分进行再分割,又得到新的目标和背景部分,如此重复下去,直至得到目标为止。对传统的Otsu方法、一维熵方法、二维熵方法的递归分割特性进行了分析比较,并根据目标的先验知识提出一种合理的递归终止准则。试验结果证明,基于递归门限分析的方法是一种行之有效的目标分割方法,分割性能优于传统方法。     

Anatomical region segmentation method from dermoscopic images of pigmented skin lesions

Melanoma tumor can cause a serious life threatening problem in humans, if left untreated for a long time without early diagnosis. For early diagnosis of melanoma, it is more significant to develop novel methods based on biophysics analyses, molecular targets recognitions, and new image analysis criteria. In this article, anatomical region segmentation and diameter identification is proposed to detect melanoma from dermoscopic images. Four main steps of the proposed system are as follows: In the first step, the preprocessing is performed to smooth the melanoma extraction process. The region segmentation is done in the second step using watershed segmentation and Sobel operator. In the third step, the postprocessing procedures like as morphological open, canny edge detection also applied to improve the region of interest. Finally, the melanoma region is identified using color symmetry features. The proposed method is tested with two data sets to prove the performance proposed method. The proposed method achieved an accuracy of 95.31% and specificity of 98.3%, which is better than other methods. Experimental results show that the effectiveness of the proposed method and illustrate viability of real-time clinical applications.     

Deeply supervised U‐Net for mass segmentation in digital mammograms

Mass detection is a critical process in the examination of mammograms. The shape and texture of the mass are key parameters used in the diagnosis of breast cancer. To recover the shape of the mass, semantic segmentation is found to be more useful rather than mere object detection (or) localization. The main challenges involved in the mass segmentation include: (a) low signal to noise ratio (b) indiscernible mass boundaries, and (c) more false positives. These problems arise due to the significant overlap in the intensities of both the normal parenchymal region and the mass region. To address these challenges, deeply supervised U‐Net model (DS U‐Net) coupled with dense conditional random fields (CRFs) is proposed. Here, the input images are preprocessed using CLAHE and a modified encoder‐decoder‐based deep learning model is used for segmentation. In general, the encoder captures the textual information of various regions in an input image, whereas the decoder recovers the spatial location of the desired region of interest. The encoder‐decoder‐based models lack the ability to recover the non‐conspicuous and spiculated mass boundaries. In the proposed work, deep supervision is integrated with a popular encoder‐decoder model (U‐Net) to improve the attention of the network toward the boundary of the suspicious regions. The final segmentation map is also created as a linear combination of the intermediate feature maps and the output feature map. The dense CRF is then used to fine‐tune the segmentation map for the recovery of definite edges. The DS U‐Net with dense CRF is evaluated on two publicly available benchmark datasets CBIS‐DDSM and INBREAST. It provides a dice score of 82.9% for CBIS‐DDSM and 79% for INBREAST.     

HSV空间下改进的直觉模糊C均值聚类猪肉图像检测

利用计算机视觉技术对畜肉分级的方法种类繁多,但由于光照因素使前期图像预处理分割目标和背景的工作难度增大。针对传统的最大类间方差法分割图像效果不佳、噪声适应能力不强的问题,以及核磁共振、高光谱成像等无损检测方法大多存在检测仪器体积大、不便于携带、成本高等问题,提出利用色调、饱和度、明度(Hue,Saturation,Value,HSV)色彩空间结合聚类方法对图像像素点进行聚类分割。在对取自自然光照环境中的猪肉图像进行分割时,所提方法相对于传统聚类分割方法分割正确率平均提高1.46%;在对人工加入了0.1椒盐噪声和0.2椒盐噪声的图像进行分割时,该方法相对于传统方法表现出了更好的抗噪声能力,传统分割方法平均错误率分别升高了16.15%和38.28%,该方法平均错误率仅升高了1.57%和1.49%。该方法具有良好的分割准确率和噪声鲁棒性,提高了目标区域的检测精度,减少了图像预处理阶段的信息丢失,提高了畜肉分级方法的质量。     

An efficient approach for brain image (tissue) compression based on the position of the brain tumor

Medical image processing plays an important role in brain tissue detection and segmentation. In this paper, a computer aided detection of brain tissue compression based on the estimation of the location of the brain tumor. The proposed system detects and segments the brain tissues and brain tumor using mathematical morphological operations. Further, the brain tissue with tumor is compressed using lossless compression technique and the brain tissue without tumor is compressed using lossy compression technique. The proposed method achieves 96.46% sensitivity, 99.20% specificity and 98.73% accuracy for the segmentation of white matter regions from the brain. The proposed method achieves 98.16% sensitivity, 99.36% specificity and 98.78% accuracy for the segmentation of cerebrospinal fluid (CSF) regions from the brain and also achieves 93.07% sensitivity, 98.79% specificity and 97.63% accuracy for the segmentation of grey matter regions from the brain. This paper focus the brain tissue compression based on the location of brain tumor. The grey matter of the brain is applied to lossless compression due to the presence of the tumor in grey matter of the brain. The proposed system achieves 29.23% of compression ratio for compressing the grey matter of the brain region. The white matter and CSF regions of the brain are applied to lossy compression due to the non‐presence of the tumor. The proposed system achieves 39.13% of compression ratio for compressing the white matter and also achieves 37.5% of compression ratio for compressing the CSF tissue. © 2016 Wiley Periodicals, Inc. Int J Imaging Syst Technol, 26, 237–242, 2016