Deeplab_v3_plus-net for Image Semantic Segmentation with Channel Compression

Abstract

The task of semantic segmentation is to correctly classify every pixel of one image. Benefit from the full convolutional neural network (FCN), the image segmentation task has step into a new stage. Since Google has shown its exploration of semantic segmentation, and proposes EncoderDecoder algorithm with Atrous Separable Convolution (Deeplab_v3_plus) method for enhancing the performance of image segmentation. Following the previous work, we will lose a great deal of detail information in the process of downsampling. Therefore, we propose to supplement the compressed channel information into the different levels feature map to compensate the details lost in the process of down sampling. This paper introduces how do we compress channel information and channel information fusion. We present a Channel Compression-Encoder-Decoder with Atrous Separable Convolution Net (CC-Deeplab_v3_plus), Eventually our CC-Deeplab_v3_plus relative to Deeplab_v3_plus in the PASCAL VOC2012 and own pest datasets get 1% Mean Intersection Over Union (MIUO) boost.