Visual media has become one of the most potent means of conveying opinions or sentiments on the web. Millions of photos are being uploaded by the people on famous social networking sites for expressing themselves. The area of visual sentiment analysis is abstract in nature due to the high level of biasing in the human recognition process. This work proposes a residual attention-based deep learning network (RA-DLNet), which examines the problem of visual sentiment analysis. We aim to learn the spatial hierarchies of image features using CNN. Since the local regions in the image convey significant sentiments, we apply residual attention model, which focuses on crucial sentiment-rich, local regions in the image. The significant contribution of this work also includes an exhaustive analysis of seven popular CNN-based architectures such as VGG-16, VGG-19, Inception-Resnet-V2, Inception-V3, ResNet-50, Xception, and NASNet. The impact of fine-tuning on these CNN variants is demonstrated in visual sentiment analysis domain. The extensive experiments on eight popular benchmark data sets are conducted and the performance is measured in terms of accuracy. The comparison of accuracy with similar state-of-the-art exhibits the superiority of the proposed work.

中文翻译：

---

用于视觉情感分析的 RA-DLNet 深度学习架构

视觉媒体已成为在网络上传达意见或情绪的最有效手段之一。人们在著名的社交网站上上传了数百万张照片来表达自己。由于人类识别过程中的高度偏见，视觉情感分析领域本质上是抽象的。这项工作提出了一个基于残余注意力的深度学习网络（RA-DLNet），它检查了视觉情感分析的问题。我们的目标是使用 CNN 学习图像特征的空间层次结构。由于图像中的局部区域传达了重要的情绪，我们应用了残差注意力模型，该模型专注于图像中关键的富含情感的局部区域。这项工作的重大贡献还包括对七种流行的基于 CNN 的架构（如 VGG-16、VGG-19、Inception-Resnet-V2、Inception-V3、ResNet-50、Xception 和 NASNet。微调对这些 CNN 变体的影响在视觉情感分析领域得到了证明。在八个流行的基准数据集上进行了广泛的实验，并根据准确性来衡量性能。精度与类似的最新技术的比较展示了所提出的工作的优越性。