Lacunarity Pooling Layers for Plant Image Classification using Texture Analysis

TL;DR

This work introduces lacunarity pooling layers to capture spatial texture heterogeneity in CNN feature maps for agricultural image classification. By implementing base lacunarity, differential box counting lacunarity, and multi-scale lacunarity—with integration into frozen backbones and fusion with GAP features—the approach enriches texture representation. Experiments on PlantVillage, LeavesTex1200, and DeepWeeds across ConvNeXt, ResNet18, and DenseNet161 demonstrate consistent improvements, with multi-scale lacunarity yielding the strongest gains and clearer class separability in embeddings. The technique offers a practical, parameter-conscious way to improve texture-aware vision models for plant disease and species identification, with potential extensions to other modalities and tasks.

Abstract

Pooling layers (e.g., max and average) may overlook important information encoded in the spatial arrangement of pixel intensity and/or feature values. We propose a novel lacunarity pooling layer that aims to capture the spatial heterogeneity of the feature maps by evaluating the variability within local windows. The layer operates at multiple scales, allowing the network to adaptively learn hierarchical features. The lacunarity pooling layer can be seamlessly integrated into any artificial neural network architecture. Experimental results demonstrate the layer's effectiveness in capturing intricate spatial patterns, leading to improved feature extraction capabilities. The proposed approach holds promise in various domains, especially in agricultural image analysis tasks. This work contributes to the evolving landscape of artificial neural network architectures by introducing a novel pooling layer that enriches the representation of spatial features. Our code is publicly available.

Figures (7)

• Figure 1: Fractal patterns exhibit an increase in lacunarity from left to right, indicating a rise in irregular gaps towards the right.
• Figure 2: Overview of the proposed fusion model. Images are passed into a frozen pre-trained network and the feature maps from the final convolution layers are passed into the lacunarity pooling layer. The lacunarity features maps are then multiplied by the feature maps from the average pooling layer for enhanced analysis.
• Figure 3: Illustration of differential box counting. The input image is divided into window sizes w x w, and further subdivided into r x r x r sized cubes stacked up until it covers the grayscale intensity of the local window. Here, u represents the lowest grayscale value and v represents the highest grayscale value.
• Figure 4: Multi-scale lacunarity feature approach is shown. Each input feature map is computed at different scales using a Gaussian pyramid. For each scale, the lacunarity feature map is computed and each scale is then upsampled to the same spatial resolution before concatenation.
• Figure 5: Example results of DeepWeeds dataset across pooling layers for the Resnet18 model. EigenCAM muhammad2020eigen was used to create the class activation maps to show which areas of the image each model focused on. As we can see in this example, the multi-scale lacunarity pooling layer captures useful information across varying spatial levels of the image.