Page 56 - Read Online
P. 56
Page 6 of 17 Hu et al. J. Mater. Inf. 2025, 5, 44 https://dx.doi.org/10.20517/jmi.2025.21
Figure 2. Flowchart: object detection process utilizing self-supervised learning.
Figure 3. Model architecture. (A) SimSiam model; (B) Faster R-CNN model. R-CNN: Region-based convolutional neural network.
(1)
The learning rate follows a cosine decay schedule, which is mathematically expressed as:
(2)
where t is the current epoch, and T is the total number of epochs. Additionally, the optimizer includes a
weight decay of 0.0001 and a momentum of 0.90. The default batch size is 512 and batch normalization
(BN) is implemented.
(2) Projection MLP: Each fully connected layer within the projected MLP component of the coding network
is succeeded by BN. The fully connected output layer does not use rectified linear unit (ReLU) activation.
The hidden layer of the fully connected network has a dimension of 2,048 and the MLP comprises three
fully connected layers.
