评级
C: 精度太低,没有开源code
可以借鉴的
An aerial scene is often flooded with many key objects or contains many objects irrelevant to the scene label, adding more difficulties to aerial image understanding.
fusing shallower features such as spatial and structural features can make features more discriminative [15], [16].
Current pooling operators in CNN are predefined and usually select the maximum or mean feature value, confining the local semantic representation ability [17], [18]. Trainable pooling can assign different weights to different local regions, and has been recently investigated in visual tasks [17], [19]
Some former solutions fuse multilayer features, but are not end-to-end [6], [7].
Later, our proposed dense connected convolutional neural network (DCCNN) demonstrated that using a dense connection structure to retain shallower features outperforms many state-of-the-art methods with much fewer parameters
简介
实验室 Qi Bi 师兄的grsl ,给的精度可够低的