Extract local visual patterns from images
Image feature extraction
Keras · Class
layers.Conv2D
Conv2D applies a learnable two-dimensional convolution across height and width. It produces a configurable number of output channels and can add a bias and activation.
How it works
Conv2D applies learnable two-dimensional convolution windows across an input's height and width. It produces the requested number of output channels, optionally adds a bias, and can apply an activation function; padding, strides, dilation, and groups control how the windows operate.
Syntax
keras.layers.Conv2D(
filters,
kernel_size,
strides=(1, 1),
padding="valid",
data_format=None,
dilation_rate=(1, 1),
groups=1,
activation=None,
use_bias=True,
kernel_initializer="glorot_uniform",
bias_initializer="zeros",
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
bias_constraint=None,
**kwargs
)Parameters
| Name | Type | Description |
|---|---|---|
filters | int | Integer; number of convolution filters and output channels. |
kernel_size | int or tuple/list of 2 integers | Convolution-window size as one integer or two integers. |
strides | int or tuple/list of 2 integers | Convolution stride along height and width. |
padding | string | Padding mode: valid or same. |
data_format | string or None | Input channel layout: channels_last or channels_first. |
dilation_rate | int or tuple/list of 2 integers | Dilation factors along the two spatial dimensions. |
groups | int | Number of groups for grouped convolution. |
activation | string or callable or None | Optional activation function. |
use_bias | bool | Whether to include a bias vector. |
kernel_initializer | initializer | Initializer for the convolution kernel. |
bias_initializer | initializer | Initializer for the bias vector. |
kernel_regularizer | regularizer or None | Regularizer for the kernel. |
Returns
A 4D tensor representing activation(conv2d(inputs, kernel) + bias).
Key points
- Conv2D performs convolution over two spatial dimensions: height and width.
- filters controls the number of output channels produced by the layer.
- kernel_size accepts one integer or two integers describing the window dimensions.
- padding supports valid and same, without case sensitivity.
- use_bias determines whether a trainable bias vector is added after convolution.
- activation optionally transforms the convolution result after bias addition.
- The output is a rank-4 tensor representing the activated convolution, with bias when enabled.
Common mistakes
- Do not combine strides greater than 1 with dilation_rate greater than 1.
- Same padding preserves spatial dimensions only when the stride is 1.
Example: Basic Conv2D model
import keras
from keras import layers
inputs = keras.Input(shape=(28, 28, 1))
x = layers.Conv2D(32, 3, padding="same", activation="relu")(inputs)
outputs = layers.Conv2D(10, 3, activation="relu")(x)
model = keras.Model(inputs, outputs)
model.summary()Creates a model with two 2D convolution layers for single-channel images.
When to use it
Build spatial feature hierarchies
Classification backbones
Reduce channel connectivity with grouped convolution
Grouped architectures
Apply the same convolution layer across time with TimeDistributed
Sequence-like image data
Interview questions
QWhat does Conv2D compute?
Conv2D slides a learned kernel over the input's two spatial dimensions and returns a tensor with one output channel per filter. With use_bias enabled, it adds a bias and then applies the optional activation.
Follow-up: How does padding affect spatial dimensions? With valid padding there is no padding; same padding pads the edges and preserves each spatial dimension when the stride is 1.
QWhat are filters and kernel_size?
filters specifies the output-space size, meaning how many convolution filters are used. kernel_size specifies the two-dimensional convolution-window size, and it may be one integer or a pair of integers.
Follow-up: What determines the output height and width? The kernel size, strides, dilation rate, and padding determine the resulting spatial dimensions.
QWhat input layouts does Conv2D support?
The layer accepts rank-4 inputs. With channels_last, the shape is (batch, height, width, channels); with channels_first, it is (batch, channels, height, width).
Follow-up: How do groups change the operation? The input channels are split into groups, each filter group operates on its corresponding input group, and the group divisibility requirements must hold.
Related
Conv1DConv3DSeparableConv2DDepthwiseConv2DConv2DTransposeMaxPooling2D