Fig. 2

The fifth layer’s feature map is initially subjected to a 1\(\times\)1 convolutional operation to modify the channel dimension (set to 256 in this chapter), resulting in a new feature map termed M5. M5 is subsequently upsampled using the nearest-neighbor interpolation method to match the dimensions of the fourth layer’s feature map. The fourth layer’s feature map, altered to match the channel dimension, is then element-wise added to M5 at corresponding positions, yielding a novel fourth layer feature map denoted as M4. This was repeated twice to get M3 and M2. The M layer feature map then went through 3\(\times\)3 convolution (to reduce the aliasing effect caused by the nearest neighbor interpolation, and the surrounding numbers were all the same). Finally, P2, P3, P4, and P5 layers were obtained