Detecting the correct orientation of an image is an important part of computer vision and the image processing pipeline. To determine the orientation of abstract paintings, as a special image type with ambiguous content, is difficult. There are several problems in the current orientation detection research: one is the use of a great deal of low-level image features for classification; the second is that the input image size is typically required to be consistent when using a deep learning model. To solve these problems, we propose a multi-scale and multi-layer feature fusion Net (MMFF-Net). First, local binary patterns (LBP) were used to generate three LBP-RGB feature maps with different scales in RGB mode, which could express the rotation characteristics of the image well; secondly, a neural network model based on AlexNet and spatial pyramid pooling (SPP) was constructed. Two data sets were selected to compare the training effect of different model parameters and model structures. Compared with the experimental results of several models, the proposed model effectively improved the accuracy of the correct orientation detection of abstract painting images.

检测图像的正确方向是计算机视觉和图像处理管道的重要组成部分。抽象绘画作为一种内容模糊的特殊图像类型，确定其方向是困难的。目前的方向检测研究存在几个问题：一是利用大量的低级图像特征进行分类；第二个是在使用深度学习模型时，输入图像大小通常需要保持一致。为了解决这些问题，我们提出了一种多尺度多层特征融合网络（MMFF-Net）。首先，使用局部二值模式（LBP）在RGB模式下生成三个不同尺度的LBP-RGB特征图，可以很好地表达图像的旋转特征；第二，构建了基于AlexNet和空间金字塔池化（SPP）的神经网络模型。选取两个数据集比较不同模型参数和模型结构的训练效果。与几种模型的实验结果相比，该模型有效提高了抽象绘画图像正确方向检测的准确性。