一种用于车辆图像分割的MSSA-UNet模型

A MSSA-UNet model for vehicle image segmentation

针对实际交通场景下的车辆图像分割方法存在模糊、效果差的问题,本文以UNet神经网络模型为基础,提出了一种融合多尺度模块和空间注意力机制的MSSA-UNet模型。在编解码阶段,采用空洞卷积构建多尺度模块,改善卷积层感受野大小受限的同时输出包含多尺度的特征信息。在上采样前,引入空间注意力机制来弥补采样过程中的局部信息丢失问题,提高特征还原能力。结合交叉熵损失与Dice损失,优化网络学习和训练过程,提高模型的分割精度。实验结果表明,本文提出的MSSA-UNet模型对于车辆图像分割任务在IoU评价指标达到83.48%,较改进前准确度提升了2.28%,模型预测值和真实值更接近,分割效果更好,有效提升了模型的分割性能。

In view of the problems of ambiguous and poor effect of vehicle image segmentation methods in actual traffic scenarios, this paper proposes a MSSA-UNet model that integrates multi-scale modules and spatial attention mechanism based on the UNet neural network model. In the encoding and decoding stage, dilated convolution is used to build a multi-scale module to improve the limited receptive field size of the convolutional layer while the output contains multi-scale feature information. Before up-sampling, a spatial attention mechanism is introduced to compensate for the problem of local information loss during the sampling process and improve the feature restoration ability. Combined with cross entropy loss and Dice loss, the network learning and training process is optimized, and the segmentation accuracy of the model is improved. The experimental results show that the MSSA-UNet model proposed in this paper achieves 83.48% in the IoU evaluation index for vehicle image segmentation tasks, which is 2.28% higher than the accuracy before improvement, the predicted value of the model is closer to the real value, and the segmentation effect is better, which effectively improves the segmentation performance of the model.