Table 4 Composition of different layers in CNNs.

Convolutional layer

The convolutional layer modifies the fully connected structure by introducing local connections between neurons and local regions of input data. Such connections are called receptive fields of neurons, and hyperparameters determine their size. In the depth dimension, the size of the receptive field remains consistent with the input dimensions. Neurons with parameters and receptive fields are referred to as convolutional kernels, which play a crucial role in the convolutional layer²⁹

Pooling layer

The pooling layer operates under the requirement of reducing the data volume size and is usually inserted periodically between convolutional layers to reduce the number of parameters. It achieves data reduction by performing pooling operations within local regions. This operation helps extract essential features and retain key information, thereby achieving efficient data representation. The introduction of pooling layers is of great significance for the performance and computational efficiency of the overall network and plays a critical role in the architectural design of CNNs³⁰

Fully connected layer

After extracting sufficient image features, the final output can be mapped to a linearly separable space using fully connected layers. The fully connected layer is similar to the fully connected layer in the Multi-Layer Perceptron (MLP), where the neurons are fully connected to all the data from the previous layer³¹. This design allows the output layer to learn and process image features comprehensively