zea.models.deeplabv3

DeepLabV3+ architecture for multi-class segmentation. For more details see https://arxiv.org/abs/1802.02611.

Functions

DeeplabV3Plus(image_shape, num_classes[, ...])

Build DeepLabV3+ model for semantic segmentation.

DilatedSpatialPyramidPooling(dspp_input)

Implement Atrous Spatial Pyramid Pooling (ASPP) module.

convolution_block(block_input[, ...])

Create a convolution block with batch normalization and ReLU activation.
zea.models.deeplabv3.DeeplabV3Plus(image_shape, num_classes, pretrained_weights=None)[source]

Build DeepLabV3+ model for semantic segmentation.

DeepLabV3+ combines the benefits of spatial pyramid pooling and encoder-decoder architecture. It uses a ResNet50 backbone as encoder, ASPP for multi-scale feature extraction, and a simple decoder for recovering spatial details.

Architecture: 1. Encoder: ResNet50 backbone with atrous convolutions 2. ASPP: Multi-scale feature extraction 3. Decoder: Simple decoder with skip connections 4. Output: Final segmentation prediction

Reference: https://arxiv.org/abs/1802.02611

Parameters:

  • image_shape (tuple) – Input image shape as (height, width, channels)

  • num_classes (int) – Number of output classes for segmentation

  • pretrained_weights (str, optional) – Pretrained weights for ResNet50 backbone. Defaults to None.

Returns:

Complete DeepLabV3+ model

Return type:

keras.Model

zea.models.deeplabv3.DilatedSpatialPyramidPooling(dspp_input)[source]

Implement Atrous Spatial Pyramid Pooling (ASPP) module.

ASPP captures multi-scale context by applying parallel atrous convolutions with different dilation rates. This helps the model understand objects at multiple scales.

The module consists of: - Global average pooling branch - 1x1 convolution branch - 3x3 convolutions with dilation rates 6, 12, and 18

Reference: https://arxiv.org/abs/1706.05587

Parameters:

dspp_input (Tensor) – Input feature tensor from encoder

Returns:

Multi-scale feature representation

Return type:

Tensor

zea.models.deeplabv3.convolution_block(block_input, num_filters=256, kernel_size=3, dilation_rate=1, use_bias=False)[source]

Create a convolution block with batch normalization and ReLU activation.

This is a standard building block used throughout the DeepLabV3+ architecture, consisting of Conv2D -> BatchNormalization -> ReLU.

Parameters:

  • block_input (Tensor) – Input tensor to the convolution block

  • num_filters (int) – Number of output filters/channels. Defaults to 256.

  • kernel_size (int) – Size of the convolution kernel. Defaults to 3.

  • dilation_rate (int) – Dilation rate for dilated convolution. Defaults to 1.

  • use_bias (bool) – Whether to use bias in the convolution layer. Defaults to False.

Returns:

Output tensor after convolution, batch normalization, and ReLU

Return type:

Tensor