Which statement about a convolutional neural network is incorrect?

The statement that a convolutional kernel cannot be used to extract global features of images is indeed incorrect because convolutional kernels, or filters, are designed to extract features from images at various levels of hierarchy, including global patterns. Convolutional layers in a convolutional neural network (CNN) apply these kernels across the entire input image (or relevant portions through the concept of a sliding window), allowing the network to recognize broad representations and patterns in the data.

While individual kernels are more effective at capturing localized features (like edges or textures), deeper layers of the network, which consist of a larger receptive field, can indeed capture more global features by combining the outputs from earlier layers. Thus, the notion that convolutional kernels cannot contribute to identifying global features is a misunderstanding of how feature extraction occurs in CNN architectures.

In contrast, the other options accurately describe aspects of convolutional neural networks. They correctly identify the composition of these networks, the types of pooling layers available, and the operational mechanism of image convolution via window scanning.