融合残差连接的图像语义分割方法

Semantic Segmentation of Residual Connection

摘要:

由于传统SegNet模型在采样过程中产生了大量信息损失，导致图像语义分割精度较低，为此提出了一种融合残差连接的新型编-解码器网络结构：文中引入了多残差连接策略，更为全面地保留了多尺度图像中包含的大量细节信息，降低还原降采样所带来的信息损失；为进一步加速网络训练的收敛效率，改善样本的不平衡问题，设计了一种带平衡因子的交叉熵损失函数，对正负样本不平衡现象予以针对性的优化，使得模型的训练更加高效；实验表明该方法较好地解决了语义分割中信息损失以及分割不准确的问题，与SegNet相比，本网络在Cityscapes数据集上进行精细标注的mIoU值提高了约13%。

Abstract:

Due to the large amount of information loss generated by the traditional SegNet model during the sampling process, the accuracy of semantic segmentation is low. Therefore, a new encoder-decoder network structure with residual connection is proposed. The multi-residual connection strategy is introduced to retain a large number of detailed information contained in multi-scale images more comprehensively, and reduce the information loss caused by decimation. In order to further accelerate the convergence efficiency of network training and improve the imbalance problem of samples, a cross-entropy loss function with balance factor is designed, and the imbalance phenomenon of positive and negative samples is optimized in a targeted manner, so that the training of the model is more efficient. Experiment shows that this method solves the problems of information loss and inaccurate segmentation in semantic segmentation, and compared with SegNet, the mIoU value of fine labeling on Cityscapes dataset is increased by about 13%.