基于非特定类别图像前景主体分割的深度学习算法研究

Study on Deep Learning Algorithm for Foreground Subject Segmentation of Non-specific Category Images

通过在Mobile-Unet网络的基础上加入SENet通道注意力机制来改进图像前景主体分割算法。改进后的算法引入深度可分离卷积来减小模型参数量,同时利用跳跃连接和多尺度特征融合来提高模型的分割精度。在训练过程中,采用了带空洞卷积的空间金字塔池化模块来增加感受野,提高模型对于大尺度物体的识别能力。实验结果表明,改进后的算法在PASCAL VOC2012数据集上达到了96%的MIOU(Modular Input/Output Unit)分割精度,准确率达到了0.971,优于现有的多种图像分割算法,例如FCN全卷积神经网络算法。在速度方面,模型对于每张图片的处理时间为1.7~2.5 s,改进后的算法相对于传统的全卷积神经网络具有更快的推理速度,适合于在移动设备上实现实时图像分割。通过对比实验,比较了改进前和改进后的Mobile-Unet模型以及FCN模型对于明亮条件下和昏暗条件下图像前景主体分割的效果,并得出了改进后的Mobile-Unet模型具有最好效果的结论。最终进行算法的部署,设计了GUI可视化操作界面,并生成.exe可执行文件。

By incorporating SENet channel attention mechanism on the basis of Mobile Unet network,the image foreground subject se-gmentation algorithm is improved.The algorithm introduces deep separable convolution to reduce the number of model parameters,while utilizing skip connections and multi-scale feature fusion to improve the segmentation accuracy of the model.Du-ring the training process,a spatial pyramid pooling module with hollow convolution is used to increase the receptive field and improve the model’s recognition ability for large-scale objects.Experimental results show that the improved algorithm achieves 96%MIOU(Modular Input/Output Un-it)segmentation accuracy on the PASCAL VOC2012 dataset,with an accuracy rate of 0.971,which is superior to various existing image segmentation algorithms,such as the FCN fully convolutional neural network algorithm.In terms of speed,the processing time of the model for each image is between 1.7 s and 2.5 s.The improved algorithm has a faster inference speed compared to traditional fully convolutional neural networks,making it suitable for real-time image segmentation on mobile devices.Through comparative experiments,the effectiveness of the Mobile Unet models before and after the improvement,as well as the FCN model,in foreground subject segmentation of images under bright and dim conditions is compared,and the conclusion is drawn that the improved Mobile Unet model has the best performance.Finally,the algorithm is deployed,a GUI visualization operation interface is designed,and an.exe executable file is generated.