基于自适应多尺度与轮廓梯度的遥感图像分割网络

Remote Sensing Image Segmentation Network Based on Adaptive Multiscale and Contour Gradient

遥感图像分割算法易受环境因素干扰,如物体遮挡、光照不均匀等。现有的深度学习遥感图像语义分割方法通常采取端到端的编解码结构,但针对相似度较高物体的结构和轮廓,仍存在分割不准确的问题。为了提高算法鲁棒性、分类准确率,提出一种基于轮廓梯度学习的深度卷积神经网络遥感图像语义分割算法。为了提高预测特征图的质量,首先基于SegNet模型,提出自适应注意力的多通道多尺度特征融合网络(D-MMA Net),其中D-MA block采用基于注意力的自适应多尺度模块,根据学习到的权重自适应地对不同尺度特征进行提取,以获得更多有效的高级语义特征。为进一步细化提取物体的边界,基于Sobel边缘检测算子原理提出可学习的轮廓提取模块。最后将轮廓信息与多尺度语义特征相结合,以增强对图像空间分辨率的鲁棒性。实验结果表明,所提算法提高分割的准确率,对于不规则物体边界,能有良好的分割效果。

Remote sensing image segmentation algorithms are susceptible to interference from environmental factors,such as object occlusion and uneven illumination.Existing deep learning remote sensing image semantic segmentation methods usually adopt an end-to-end codec structure.However,they still suffer from inaccurate segmentation for the structure and contours of high similarity objects.Therefore,to improve the algorithm robustness and classification accuracy,a deep convolutional neural network remote sensing image semantic segmentation algorithm based on contour gradient learning is proposed.To improve the quality of the predicted feature maps,the adaptive attention-based multichannel multiscale feature fusion network(D-MMA Net) is proposed based on the SegNet model network.The DMA block uses an attention-based adaptive multiscale module to adaptively extract different scale features according to the learned weights to obtain more effective high level semantic features.To further refine the extracted object boundaries,the contour extraction module,a learnable contour extraction module,is proposed based on the principle of the Sobel edge detection operator.Finally,the contour information is combined with multi-scale semantic features to enhance the robustness of the spatial resolution of the image.The experimental results show that the proposed method improves the segmentation accuracy and produces good segmentation results for irregular object boundaries.