School of Software, East China Jiaotong University
National Natural Science Foundation of China,the Humanity and Social Science Fund of Ministry of Education of China
材质图像识别具备广阔的应用前景,如衣物识别、机器人拾取、工业检测等.受光照强度、拍摄角度等影响, 材质图像易发生变化. 而挖掘鲁棒、高效的图像特征是应对该变化的关键. 提出SECF2模型:抽取SENet中具有良好互补性的异构层特征; 改进聚类典型相关性分析模型, 实现异构层特征融合, 生成刻画材质图像的深层视觉语义, 它是一种判别性更强且鲁棒的新特征; 采用深层视觉语义训练分类模型并执行集成学习, 完成材质图像识别. 实验表明: SECF2模型在两个材质图像数据集上都有效, 其中Fabric上的识别精准度较最强基线提升8.85%.SECF2模型还具备较强通用性, 在图像情感分析基准数据集上取得了优异的表现.此外, SECF2仅需两个特征和一次融合, 模型复杂度降低且实时效率优良.
Material image recognition has broad application prospects, such as clothing recognition, automatic picking by a robot, industrial detection, etc. Owing to the influence of light intensity and camera angle, material image is easy to change. Mining robust and effective image features is one of the most important factors to handle this change. To address this problem, a novel model called SECF2 is proposed. Heterogeneous-layer features in the SENet model are extracted respectively. These heterogeneous-layer features can complement each other well. Then the traditional Cluster-CCA model is modified to complete the Feature Fusion of these heterogeneous-layer features. As a significant result, more discriminant and robust deep-level visual semantics is obtained after feature fusion to better characterize the material image. Finally, the deep-level visual semantics is employed to train classification models, and a voting-based ensemble learning strategy is designed to further boost the final classification performance. Experimental results demonstrate that the SECF2 model is effective on two material image datasets. Especially, compared with the best baseline, it obtains a 8.85% performance improvement in the Fabric dataset. In addition, SECF2 model has strong versatility. In the image sentiment analysis task, it has achieved excellent performance. Moreover, the proposed SECF2 model only needs two features and one time feature fusion. Hence, the model obtains better real-time efficiency.