The variation in pixel values between different objects during semantic segmentation of images leads to the loss of local image details in existing network models. An image semantic segmentation method (DECANet) is proposed to solve this problem. First, a channel attention network module is introduced to improve network clarity by modeling the dependencies of all channels, selectively learning and reinforcing channel features, and extracting useful information to suppress useless data. Second, using an improved atrous space pyramidal pooling (ASPP) structure, the extracted image convolutional features are multiscale fused to reduce the loss of image detail information, and the semantic pixel location information is extracted without increasing the weight parameters to speed up the model's convergence. Finally, the mean intersection over union of the proposed method reaches 81.08% and 76% on PASCAL VOC2012 and Cityscapes datasets, respectively. The detection performance of the DECANet is superior to the existing state-of-the-art network models, which can effectively capture local detail information and reduce image semantic pixel classification errors.