为解决工厂流水线上不同种类动态物品的快速精准抓取问题，提出一种两阶段动态多物品定位抓取方法。第1阶段采用所提多尺度上下文感知的单分支融合语义分割网络获取目标物品的掩码区域：首先特征提取网络采用单分支结构，在保证提取丰富的空间信息和高层语义信息的同时，减小网络参数量；随后特征融合网络通过双边引导特征融合模块增强空间信息和语义信息的表达能力；最后设计特征增强网络，通过特征辅助收敛模块嵌入浅层和深层网络中，加快网络收敛速度。第2阶段采用基于轮廓点检测的快速位姿估计策略在掩码区域预测最佳抓取点位姿。在自建数据集上的测试及流水线平台抓取实验结果表明，所提方法能实时检测和预测物品抓取点位姿，精准完成物品抓取，其分割精度、预测时间和抓取成功率均优于对比方法。

针对自然场景下混凝土裂缝背景复杂、纹理多变所造成的裂缝分割不准确问题，提出了一种基于特征增强整体嵌套网络裂缝检测算法。首先，在整体嵌套网络，一种深度学习边缘检测网络的基础上，采用多尺度监督机制融合不同尺度混凝土裂缝的预测结果，增强网络对混凝土裂缝线性拓扑结构的表达能力；然后，采用一种卷积-反卷积特征融合模块将混凝土裂缝的反卷积深层语义特征和卷积浅层细节特征有效融合，深层语义特征可以减少复杂背景干扰并提高模糊裂缝区域特征响应，浅层特征可以提高裂缝细节表达能力，进而提高裂缝特征质量；最后，提出了混合空洞卷积边界细化模块，该模块利用残差网络和空洞卷积组细化裂缝边界，从而提高裂缝分割准确性。在数据集Bridge_Crack_Image_Data和Crack Forest Dataset上进行了测试，所提算法的准确率分别达到了92.1%和91.6%，F1分数分别达到了80.2%和91.1%。实验结果表明，所提算法在复杂自然环境下可获得稳定而准确的分割结果，具有较强的泛化性。

提出了一种基于多引导滤波器的单幅图像超分辨率方法。首先，该方法通过大量的自然图像建立高低分辨率图像块样本训练库，并通过聚类算法将具有相似性质的高低分辨率样本块进行聚类；其次，将输入低分辨率图像进行重叠分块，并在样本库中搜索最近邻的高低分辨率样本聚类；再次，将输入低分辨率图像块作为输入图像，与样本库中最近邻的低分辨率聚类样本作为引导图像，运用本文提出的多引导滤波器计算引导滤波器的参数；最后，利用样本库中最近邻的高分辨率聚类样本和引导滤波器的参数，通过多引导滤波器就可以重构高分辨率图像。实验结果表明，本文算法不仅能很好地重构图像的高频细节，还能很好地恢复图像的纹理特征。

Fusion of multiple instances within a modality for biometric verification performance improvement has received considerable attention. In this letter, we present an iris recognition method based on multi-instance fusion, which combines the left and right irises of an individual at the matching score level. When fusing, a novel fusion strategy using minimax probability machine (MPM) is applied to generate a fused score for the final decision. The experimental results on CASIA and UBIRIS databases show that the proposed method can bring obvious performance improvement compared with the single-instance method. The comparison among different fusion strategies demonstrates the superiority of the fusion strategy based on MPM.

Reflectance model is a basic concept in computer vision. Some existing models combining the classical diffuse reflectance model and those for surfaces containing specular components can approximately describe real reflectance. But the ratio of diffuse and specular reflection decided manually has no clear meaning. We propose a new polynomial hybrid reflectance model. The reflectance map equation with a known shape (for example cylinder) as a sample is used to estimate parameters of the proposed reflectance model by least square regression algorithm. Then the reflectance parameters for surfaces of the same class of materials can be determined. Experiments are performed for a metal surface. The synthesis images produced by the proposed method and existing ones are compared with the real acquired image, and the results show that the proposed reflectance model is suitable for describing real reflectance.

Shape-from-shading (SFS) is to reconstruct three-dimensional (3D) shape from a single gray image, which is an important problem in computer vision. We propose a novel SFS method based on hybrid reflection model which contains both diffuse reflectance and specular reflectance. The intensity gradient of image is in the direction that the shape of surface changes most, so we use directional derivative of the reflectance map as parts of objective function. When discrete characteristic of digital images is considered, finite difference approximates differential operator. So the reflectance map equation described by a partial differential equation (PDE) turns into an algebra equation about the unknown surface height correspondingly. Using iterative numeric computation, a new SFS method is gained. Experiments on synthesis and real images show that the proposed SFS method is accurate and fast.