基于神经网络的图像弱监督语义分割算法

摘要

Category activating feature map algorithm is an algorithm which is able to explore the feature map according to specific category and train it by weakly supervised sample,and the semantic information extracted could be provided to other detection or location mission.Thus, an image semantic segmentation algorithm which is calculated by neural network is proposed, the model could be obtained by training the neural network by weakly supervised training data. This algorithm calculated the rough region by semantic segmentation which combined the feature image from neural network with the network parameter,then obtained the more accurate image semantic segmentation by passing back the semantic information.Extensive experiments show that the proposed method has competitive performance on many datasets.Furthermore,we visualize some results to explain the details of the proposed algorithm.%类别激活热度图算法是一种可以在图像中找到具体分类对应的热度图的使用弱监督样本进行训练的算法,算法提取得到的语义信息可以提供给其他的检测任务或者定位任务所使用.提出一种使用神经网络进行计算的图像语义分割的算法,仅需要使用弱监督的训练数据对神经网络进行训练,得到模型.该算法将神经网络所输出的特征图像与网络参数相结合计算得到语义分割的大致区域,再在其中使用语义信息回传的方法,从大致区域的结果中得到更为精确的图像语义分割.最后介绍了该算法在不同的数据集上进行验证的结果,并且展示了内部的实现细节.