IEEE Geoscience and Remote Sensing Magazine - December 2023 - 106

SEGMENTER [174]
The Segmenter network relies on the output embeddings
corresponding to image patches and obtains class labels
from these embeddings with a point-wise linear decoder or
a mask Transformer decoder (see Figure 9). This work considers
the tiny ViT [151].
HRNET [170]
HRNet can maintain high-resolution representations throughout
the whole process. HRNet starts from a high-resolution
convolution stream, gradually adds high-to-low-resolution
convolution streams one by one, and connects the multiresolution
streams in parallel (see Figure 10). The resulting network
consists of several stages (four in the original article), and the
nth stage contains streams corresponding to resolutions. In this
work, we use the HRNet small version.
STDC [152]
STDC is a novel and efficient structure for removing structure
redundancy. Specifically, the network gradually reduces
the dimension of feature maps and uses their aggregation
for image representation, forming the STDC network's basic
module. In the decoder, a detail aggregation module is proposed
by integrating the learning of spatial information into
low-level layers in a single-stream manner. Finally, the lowlevel
features and deep features are fused to predict the final
segmentation results (see Figure 11).
CGNet [173]
CGNet is a lightweight network for semantic segmentation.
It uses a context-guided block to learn joint features of both
local and surrounding contexts (see Figure 12). Based on
the context-guided block, CGNet can capture contextual
information in all stages of the network, specially tailored
for increasing segmentation accuracy. CGNet is also elaborately
designed to reduce the number of parameters and
save the memory footprint.
We also present a brief introduction of the NAS methods
that are employed for the study:
SqueezeNAS [169]
SqueezeNAS is considered the first proxyless hardwareaware
search that was targeted for dense semantic segmentat
ion. Using a different iable search strategy, it
Segmenter
+
+
+
+
+
1
2
7
8
9
Transformer
Encoder
Mask
Transformer
Class
Masks
Patch Embedding
Position Embedding
Patch Encoding
Input Image
(a)
Tree
Sidewall
Person
(b)
Segmentation Map
FIGURE 9. An overview of the Segmenter framework. (a) Encoder: image patches are projected to an embedding sequence and then
encoded with a transformer. (b) Decoder: a mask transformer receives the output of the encoder and class embeddings as input to predict a
segmentation mask. (Adapted from [174].)
Scalar
Product
Channel
Maps
Strided
Conv.
Upsample
Conv.
Unit
FIGURE 10. The main idea of the HRNet architecture. It consists of high-to-low resolution parallel subnetworks with multiscale fusion. The
horizontal direction indicates the depth of the network; the vertical direction indicates the scale of the feature maps. Conv.: convolution.
(Adapted from [170].)
106
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2023
Flatten and Project
Upsample and Argmax

IEEE Geoscience and Remote Sensing Magazine - December 2023

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2023