现有的移动端人像分割网络存在分割精度差、分割边缘模糊等问题。为此, 提出了一种融合注意力机制的轻量化人像分割网络。首先, 利用MobileNetV2网络提取图像特征。然后对注意力模块NLNet(Non-local neural networks)进行轻量化处理, 随后将优化过的注意力模块嵌入到四层解码网络中。利用融合注意力机制的解码网络自适应地学习有效特征, 最后通过SoftMax层得到人像分割结果图。同时改进了损失函数, 引入多损失函数(Multi-Loss), 使网络更容易收敛。解码网络融合注意力机制的方式使得轻量化网络可以在语义分割任务上取得较好的效果。实验结果表明, 模型在550张自采集的人像测试集上达到了92.29%的交并比(MeanIOU),单张图片在Inter(R) Core i5 CPU上的分割时间为0.74 s。与传统的人像分割网络相比, 研究网络的分割精度和分割速度优势明显, 适合应用于移动端设备。

Existing mobile terminal segmentation network has problems such as poor segmentation accuracy and segmentation edge blurring. To this end, a lightweight portrait segmentation network that incorporates attention mechanisms is proposed. Firstly, the image features are extracted by MobileNetV2 network. Then, the attention module NLNet (Non-local neural networks) is lightened, and then the optimized attention module is embedded in the four-layer decoding network. The decoding network with attention mechanism adaptively learns the effective features, and finally obtains the portrait segmentation result map through SoftMax. At the same time, the loss function is improved, and Multi-Loss (multi-loss function) is introduced to make the network easier to converge. The way of attention mechanisms fused in decoding networks makes the lightweight network achieve better results in the semantic segmentation task. Experimental results show that the model achieves 92.29% MeanIOU on the 550 self-collected portrait test set, and the split time of the single picture on the Inter(R) Core i5 CPU is 0.74 s. Compared with the traditional portrait segmentation network, the research network has obvious advantages in segmentation precision and segmentation speed, and is suitable for mobile devices.