关键词: 眼底图像, 中央凹检测, 全卷积网络, 金字塔池化模块

Abstract: Aiming at the problem that it was difficult to recognize the fovea of color fundus image with traditional algorithm, we proposed  a fovea automatic detection algorithm based on fully convolutional networks (FCN). Firstly, we mined global context information through the local context environment of color fundus images, and constructed an FCN model to realize local pixel-level classification. Secondly, the local pixel\|level features were extended to the global pyramid pooling module, so that the spatial statistical data could provide a better description and expression for the global context understanding, thus effectively obtain the highly differentiated global context information. Finally, the global and local features were fused to achieve accurate detection of the central fovea. The experimental results show that the algorithm improves the specificity of detection of dark fundus lesions, and provides an effective basis for the detection of serious fundus lesions.

Key words: fundus image, fovea detection, fully convolutional networks (FCN), pyramid pooling module