A convolutional layer is designed to learn local visual features from an input image. During training, small learnable filters move across the image and respond to patterns such as edges, corners, textures, and progressively more complex structures. The resulting feature maps preserve useful spatial relationships while transforming raw pixels into representations that later layers can use. Oracle documentation recognizes convolutional neural networks as suitable neural-network architectures for visual data and identifies architectures such as ResNet for processing images. Oracle Docs The convolutional layer itself does not primarily generate images or make the final classification decision. Reducing spatial dimensions is normally performed through pooling or strided operations. Therefore, detecting specific features in the input image is the correct function.

================