基于双分支特征融合的无锚框目标检测算法

针对无锚框目标检测算法CenterNet中,目标特征利用程度不高、检测结果不够准确的问题,该文提出一种双分支特征融合的改进算法。在算法中,一个分支包含了特征金字塔增强模块和特征融合模块,以对主干网络输出的多层特征进行融合处理。同时,为利用更多的高级语义信息,在另一个分支中仅对主干网络的最后一层特征进行上采样。其次,对主干网络添加了基于频率的通道注意力机制,以增强特征提取能力。最后,采用拼接和卷积操作对两个分支的特征进行融合。实验结果表明,在公开数据集PASCAL VOC上的检测精度为82.3%,比CenterNet算法提高了3.6%,在KITTI数据集上精度领先其6%,检测速度均满足实时性要求。该文提出的双分支特征融合方法将不同层的特征进行处理,更好地利用浅层特征中的空间信息和深层特征中的语义信息,提升了算法的检测性能。     

An in-depth study of tool wear monitoring technique based on image segmentation and texture analysis

We present a new micro-vision system for tool wear monitoring, which is essential for intelligent manufacturing. The tool wear area is divided into regions by a watershed transform, then subjected to automatic focusing and segmentation. The individual pixel gray values in each region are then replaced with the corresponding regional mean gray value. A hill climbing algorithm based on the sum modified laplacian (SML) focusing evaluation function is used to search the focal plane. In addition, we implement an adaptive Markov Random Field (MRF) algorithm to segment each region of tool wear. For our MRF model, the connection parameter value is adaptively determined by the connection degree between regions, which improves image acquisition of more integral tool wear areas. Our findings suggest that automatic focusing and segmentation of the tool wear area by region (within the tool wear area) enhance accuracy and robustness, and allow for real time acquisition of tool wear images. We also implement a complementary tool wear assessment procedure based on the surface texture of the workpiece. The optimal texture analysis window is determined using the entropy metric – a texture feature generated using a Gray Level Co-occurrence Matrix (GLCM). In the best texture analysis window, entropy remains monotonic as tool wear increases, demonstrating that entropy can be used effectively to monitor tool wear. Information from combined measurements of tool wear and workpiece texture can reliably be used to monitor tool wear conditions and improve monitoring success rates.     

Research on Facial Fatigue Detection of Drivers with Multi-feature Fusion

In order to solve the shortcomings of current fatigue detection methods such as low accuracy or poor real-time performance, a fatigue detection method based on multi-feature fusion is proposed. Firstly, the HOG face de-tection algorithm and KCF target tracking algorithm are integrated and deformable convolutional neural network is introduced to identify the state of extracted eyes and mouth, fast track the detected faces and extract con-tinuous and stable target faces for more efficient extraction. Then the head pose algorithm is introduced to detect the driver''s head in real time and obtain the driver''s head state information. Finally, a multi-feature fusion fatigue detection method is proposed based on the state of the eyes, mouth and head. According to the experimental results, the proposed method can detect the driver''s fatigue state in real time with high accuracy and good ro-bustness compared with the current fatigue detection algorithms.     

SIM-MFR: Spatial interactions mechanisms based multi-feature representation for background modeling

Moving object detection is frequently used as a springboard for advanced computer vision analysis in complex scenes. Nevertheless, due to unstable changes in the background, most existing background model hardly maintain superior performance. To this concern, we propose a novel pixel-level background model that has three innovations. First, we introduce K-means to directly model the spatiotemporal dependencies between pixels. These dependencies are exploited to discover static core information in the high-frequency changing spatial domain, resulting in excellent property in dynamic backgrounds. Besides, the notion of complementarity is taken as a feature selection criterion. In multi-feature model, the ability to supervise each other between features is important in the ambiguity challenges, e.g., shadow. Finally, feature models recommend each other in the update mechanism, and the diffusion rate of effective information in each feature model can be maximized by finding the best candidate feature. By virtue of this mechanism, model can be updated efficiently when large background migration occurs, e.g., PTZ. Experimental results on some standard benchmarks show that SIM-MFR can achieve promising performance compared to some state-of-the-art approaches.     

考虑特征差异的多特征工件依次定位

现有的工件定位算法对于多特征工件按照组合特征进行定位,而实际上不同特征对于定位结果的影响是不同的。考虑到工件各个特征之间的差异,在深入研究多特征的位形空间理论的基础上,参考形位误差评定问题中多基准建立的过程以及对称工件定位算法,提出了多特征工件依次定位问题和算法,按照一定准则(如特征本身精度高低)对各个特征依次进行定位。仿真试验表明该算法更符合实际。     

一种弱小目标的自适应搜索方法

针对成像跟踪系统中弱小目标搜索定位的难题,提出了一种新的自适应搜索算法,即利用全局寻优的自适应遗传算法搜索目标,以目标的多特征融合信息作为最佳定位的判断准则和适应度函数。实验结果表明,该方法不仅提高了检测精度,也改善了跟踪算法的稳定性和智能策略。     

基于序列特征融合的蛋白质可溶性预测

蛋白质可溶性在药物设计的研究中起着重要的作用,传统生物实验测试蛋白质可溶性费时费力,因此基于计算方法对可溶性进行预测成为一个重要的研究方向.针对传统可溶性预测模型不能充分表示蛋白质特征的问题,文中设计了一种基于多种蛋白质序列信息的神经网络模型PSPNet,并应用到蛋白质可溶性预测中.该模型首先使用氨基酸残基序列嵌入信息...     

Semiconormed possibility integrals and multi-feature pattern classification

The semiconormed possibility integrals are proposed as a multi-feature pattern classification model. A semiconormed possibility integral is a nonlinear integration of a function and its corresponding non-normalized possibility measures over feature space. The function of an object's feature vector represents the possibilities with uncertainty that the object belongs to a class. The uncertainty is due to the similar characteristics of objects from different classes and the distortion of the original characteristic information caused by feature data acquisition systems. The uncertainty is assessed by the non-normalized possibility measures, a possibility measure of a feature is considered as the credibility of the feature to provide reliable information for pattern classification. Integration of a function and the possibility measures effectively reduces the uncertainty and improves the pattern classification results. A pattern classification algorithm based on the semiconormed possibility integrals was used to classify a set of “ellipse data” and the well-known IRIS data, the classification results were compared with those obtained by using Bayes classifier.     

多特征融合的目标识别与提取在空间定位中的研究

提出了一种多特征融合的概念，应用颜色和形状信息以及一系列的策略实现了特征目标的识别与提取，并采用随机Hough变换技术对几何基元进行了准确的提取定位。该方法应用在交通事故的定位中得到了良好的效果。