图像语义分割基于图像场景的“语义”为目标类别中的每个像素分配标签,来区分图像中的不同种类的事物。现有基于DeepLabV3+的语义分割方法具有高计算复杂性和大内存消耗,且在提取图像特征信息时难以充分利用多尺度信息,这可能导致详细信息的丢失,降低分割精度。户外环境中的目标种类繁多、光照条件变化大以及存在遮挡,增加了场景理解和对象识别的难度。因此,文中提出改进的DeepLabV3+网络户外多目标场景分割方法,以改进的 MobileNetV2作为模型主干;将ECAnet通道注意力机制应用于低级特征,降低计算复杂性并提高目标边界清晰度;在ASPP模块之后引入了极化自注意力机制,改善特征图的空间特征表示。改进后的模型在户外数据集Standford Background Dataset上的平均交并比和平均准确度分别为69.5%和82.35%,与原始DeepLabV3+模型相比提升了5.2%和4.5%,新增模块对运行时间无显著影响,有效提高了模型的推理效率与精度。

Image semantic segmentation assigns labels to each pixel in the target category based on the “semantics” of the image scene, distinguishing different types of things in the image. Existing semantic segmentation methods based on DeepLabV3+ have high computational complexity and large memory consumption, and it is difficult to fully utilize multi-scale information when extracting image feature information, which may lead to the loss of detailed information and reduce segmentation accuracy. The outdoor environment has a variety of targets, large changes in lighting conditions, and obstructions, which increase the difficulty of scene understanding and object recognition. Therefore, an improved DeepLabV3+ network outdoor multi-target scene segmentation method is proposed, with the improved MobileNetv2 as the model backbone. The ECAnet channel attention mechanism is applied to low-level features to reduce computational complexity and improve target boundary clarity. After the ASPP module, a polarized self-attention mechanism is introduced to improve the spatial feature representation of the feature mPA. The improved model has an average intersection to union ratio (mIoU) and average accuracy (mPA) of 69.5% and 82.35% on the outdoor dataset Standford BackgroundDataset, respectively. Compared with the original DeepLabV3+model, it has improved by 5.2% and 4.5%. The addition of new modules has no significant impact on the running time, effectively improving the inference efficiency and accuracy of the model.