基于轻量级卷积神经网络的多视觉特征图像分割研究

Research on multi-visual feature image segmentation based on lightweight convolutional neural networks

研究基于轻量级卷积神经网络的多视觉特征图像分割方法,以适应资源受限环境并满足实时性需求。在Linknet网络基础上设计轻量级多视觉特征图像分割模型,以原始多视觉特征图像为编码器输入,经过初步特征提取后,多尺度特征提取模块利用不同尺度卷积核学习其颜色、纹理等特征,通道注意力模块采用压缩-激励块对不同尺度特征作重定向,利用引入深度可分离卷积的特征提取模块A学习更抽象的特征表示,解码器利用特征提取模块B、反卷积层和标准卷积层对编码器提取的特征表示作转换处理,生成包含语义信息的特征图。通过双向特征金字塔网络融合编解码器输出特征,利用Sigmoid函数获得多视觉特征图像分割结果。实验结果表明:该方法训练损失仅为0.08;可实现多视觉特征图像的精准分割,MIoU、F1-score指标分别为0.9128、0.9068;分割模型参数量、计算量、存储空间分别为6.14 MB、1.52 GMac、0.146 GB,满足轻量级要求。

feature representations.The decoder uses feature extraction module B,deconvolution layer and standard convolution layer to transform the feature representations extracted by the encoder to generate feature maps containing semantic information.The bidirectional feature pyramid network is used to fuse the encoder and decoder to output features.The Sigmoid function is used to obtain multi-visual feature image segmentation results.The experimental results show that the training loss of the method studied is only 0.08,the method can achieve accurate segmentation of multi-visual feature images with MIoU(mean intersection over union)and F1-score of 0.9128 and 0.9068,respectively,and the parameter quantity,computational complexity and storage space of the segmentation model are 6.14 MB,1.52 GMac and 0.146 GB,respectively,so the method meets the lightweight requirements.