联合多连接特征编解码与小波池化的轻量级语义分割

语义分割是当前场景理解领域的基础技术之一。现存的语义分割网络通常结构复杂、参数量大、图像特征信息损失过多和计算效率低。针对以上问题,基于编-解码器框架和离散小波变换,设计了一个联合多连接特征编解码与小波池化的轻量级语义分割网络MLWP-Net(Multi-Link Wavelet-Pooled Network),在编码阶段利用多连接策略并结合深度可分离卷积、空洞卷积和通道压缩设计了轻量级特征提取瓶颈结构,并设计了低频混合小波池化操作替代传统的下采样操作,有效降低编码过程造成的信息丢失;在解码阶段,设计了多分支并行空洞卷积解码器以融合多级特征并行实现图像分辨率的恢复。实验结果表明,MLWP-Net仅以0.74 MB的参数量在数据集Cityscapes和CamVid上分别达到74.1%和68.2%mIoU的分割精度,验证了该算法的有效性。

GPS/INS组合导航中两步自适应滤波方法

GPS/INS组合导航实际应用中存在系统模型偏差、噪声模型不确定等问题,导致卡尔曼滤波器无法实现最优滤波效果,严重时甚至导致滤波发散。渐消卡尔曼滤波器和自适应卡尔曼滤波器通过引入单渐消因子和单自适应因子可以部分解决上述问题,但是不足在于单因子只能进行整体调整,不能精确调整各个通道。针对此问题,本论文提出一种2步自适应卡尔曼滤波算法,构造基于残差协方差估计的多重渐消因子和自适应因子对各个通道精确调整,克服动态环境下跟踪性差的局限性。实验结果表明,改进后的自适应卡尔曼滤波算法可以精确调整各通道,增强系统的定位精度、跟踪性能和鲁棒性。

快速超轻量城市交通场景语义分割

近年来,随着自动驾驶的火热发展,越来越多研究者开始探索图像语义分割网络的轻量化并将其应用于道路交通场景。而目前现存的语义分割网络通常由于参数量庞大难以部署在硬件资源有限的边缘设备,针对这一问题,设计了一个由通道注意力骨干网络(CABb)和空间注意力解码器(SAD)模块构成的双注意力轻量化网络(DALNet),结合“通道-空间”双注意力机制的DALNet在图像上下文语义信息的提取和图像空间信息的恢复上都具有突出的表现。CABb主要由通道注意力瓶颈(CABt)模块组成,CABt模块采用Split策略分离特征通道并行处理多尺度的特征图,引入通道注意力机制进行通道融合,提取多尺度语义信息。SAD模块利用空间注意力机制指导解码器进行双线性插值上采样,恢复分割目标边沿以及细节信息。实验结果表明,DALNet仅凭48万的参数量在城市交通数据集Cityscapes和CamVid最高分别可达到74.1%和70.1%的交并比(mIoU)。DALNet在输入图像分辨率为512×1024的情况下,基于GTX 1080TiGPU可以获得74 frame/s的前向推理速度,远超实时语义分割所需的速度要求。

融合多尺度空洞卷积与反卷积的轻量化目标检测

深度神经网络存在目标检测速度慢、参数量大的问题,不适用于算力有限但速度要求较高的移动应用场景。为了提高目标检测的推理速度,有效权衡目标检测任务的精度与速度,文中提出了一种融合多尺度空洞卷积与反卷积的轻量化目标检测网络MDDNet。首先,基于高效的单阶段多目标检测策略设计了轻量的目标检测基础网络,并引入深度可分离卷积,以进一步减少基础网络的参数量,加快图像特征提取的速度;然后在主干网络中添加两条基于多尺度空洞卷积的特征扩展旁路,分别连接在基础网络的最末端和次末端残差层的输出端,将两条旁路的特征输出到预测层进行特征融合,以增强较低层特征图的纹理特征;并且进一步引入了多尺度反卷积模块,连接于深层特征网络层,以增大特征图尺寸,再融合具有不同尺度的上一层的浅层特征图,以获得更多的特征语义信息和细节信息,提高检测精度;最后在预测层基于K均值算法优化先验框参数,使其与目标真实框更匹配,提高目标识别的准确率。实验结果表明:MDDNet的参数量约为7.21×10^(6),平均检测精度在KITTI、Pascal VOC数据集上分别为58.7%、76.0%,推理速度在两个数据集上分别达到55和52 f/s。因此,MDDNet在参数量、检测速度和检测精度上达到了较佳的平衡,可适用于移动端的实时目标检测。

Blind source separation by weighted K-means clustering

Blind separation of sparse sources （BSSS） is discussed. The BSSS method based on the conventional K-means clustering is very fast and is also easy to implement. However, the accuracy of this method is generally not satisfactory. The contribution of the vector x（t） with different modules is theoretically proved to be unequal, and a weighted K-means clustering method is proposed on this grounds. The proposed algorithm is not only as fast as the conventional K-means clustering method, but can also achieve considerably accurate results, which is demonstrated by numerical experiments.

Chromosome G-banding in SituHybridization of RFLP Marker umc58 Linked with the Gene hm1 Dictating Helminthosporium carbonum Susceptibility1 in Maize

Two set of Andrias davidanus fertilized eggs and its seedlings were got by means of artificial reproduction. The sections of the embryo of Andrias davidianus were made by histological method. The expression of homologous sequences of Quox 1 gene has been studied by immunohistochemical method with special antibody. Results show that the expression of homologous sequences of Quox 1 gene was detected in the epidermis of blastula and gastula.The highly expression was investigated in the epidermis at the neural tube phrase and the tail bud stage.

Spacer Length Variation in Rice 5SrRNA Genes Revealed by Polymerase Chain Reaction

Polymerase chain reaction(PCR) was used to amplify 5S rRNA spacer from wild rice(Oryza rufipogon and O.nivara) and cultivated rice(indica and japonica varieties of O.sativa L).The results show that there is spacer length variation within and between species,and the typical indica and japonica varieties have their unique banding patterns of amplified 5S rRNA spacers,whereas intermediate showed no specific amplification profile of spacer regions.The 5S rRNA genes in intermediate are either identical with that of indica variety or that of japonica variety.These data suggest that the spacer length polymorphisms can be used to distinguish between closely ralated species and subspecies.

Rapid Clasification of Bacilus Isolates Using RAPD Technique

This paper describes the typing of selected Bacillus strains from environmental origin and ATCC type strains. The objective of the study was to evaluate the usefulness of PCR fingerprinting using Randomly Amplified Polymorphic DNA (RAPD) as a method for classifying members of the genus bacillus . Our experiments show that the levels of similarity of 69%～90%, 40%～60% were observed between the members of the same species and among the species of genus Bacillus , respectively. In a dendrogram showing the positions of 28 strains, these strains were clearly seperated into distinct species, 13 of which are corresponding to type strains used in this study. These results suggest that RAPD analyses is a simple and powerful method for species identification and delimitation.

Taxonomic Relationship Between Oryza minuta and O． officinalis Based on AP－PCR Analysis

The arbitrarily primed polymerase chain reaction(AP-PCR)technique was used to detect polymorphisms in genomic DNA from Oryza minuta, O. officinalis, and related rice species. The results show that the genetic distance between O.minuta and O.officinalis is much higher than that between two known species,O.sativa and O. rufipogon, showing that O.minuto and O.officinalis should be classified as two different spedes in genus Oryza,and that the AP-PCR technique can be used as a powerful tool for studying taxonomic relationships among related spedes, and even subspecies.

多尺度特征融合的道路场景语义分割

针对现有语义分割网络模型难以在参数量、推理速度和精确度中取得平衡的问题,设计了一种多尺度特征信息融合的轻量级网络模型(MIFNet)。MIFNet采用编码-解码结构,在编码部分利用分离策略和非对称卷积设计了轻量型特征提取瓶颈结构,且引入空间注意力机制与Laplace边缘检测算子组成边缘-空间融合模块,将空间信息和边缘信息进行融合得到丰富的特征信息。在解码部分引入通道注意力机制恢复特征图尺寸和细节信息完成语义分割。在Cityscapes和CamVid测试集上,MIFNet仅以0.82 M的参数量分别取得了73.1%和67.7%的分割精度,同时在单个GTX 1080Ti GPU下分别获得73.68 frame/s和85.16 frame/s的推理速度,表明该方法在参数量、推理速度和精确度3个指标上得到较好平衡,实现了轻量、快速、精准的语义分割。