【摘要】 针对现存方法对林地检测结果细节丢失严重、边缘不够精细的问题,文中提出一种基于多尺度特征融合的网络结构,使用Labelme手动对数据集进行标签,并通过翻转、加噪、颜色调整等操作增强图像。网络结构主要包括将低层空间信息和高层语义信息进行特征融合的密集跳跃连接,以及可以增大网络感受野并从多个角度对特征进行提取的空洞空间金字塔池化模块。文中利用编码器提取全局信息,通过解码器恢复图像空间分辨率,最后用分类器进行林地与非林地的分割。实验结果表明,文中算法与现有算法相比,图像检测能力有一定的提升,能较为准确地检测出林地区域。更多还原

【Abstract】 As the forestland detection results of the existing methods have the problems of serious loss of details and insufficient refinement of edges,a network structure based on multi-scale feature fusion is proposed. The Labelme is used to manually label the dataset,and the images are enhanced by flipping,adding noise,adjusting colors,etc. The network structure mainly includes dense jump connections for feature fusion of low-level spatial information and high-level semantic information,and hollow spatial pyramid pooling module which can increase the network receptive field and extract features from multiple angles. The encoder is used to extract the global information,the decoder is used to restore the image spatial resolution,and the classifier is used to divide the forestland from the non-forestland. The experimental results show that in comparison with the existing algorithms,the algorithm proposed in this paper has a certain improvement in the aspect of the image detection ability,and can detect the forestland area more accurately.更多还原