Tomato ripeness classification with weight-based and texture-adaptive pooling in CNNs

Abstract

This research introduces a weight-based mixed pooling and adaptive fuzzy pooling mechanism based on feature map texture analysis for attention mechanism Squeeze-and-Excitation Networks (SE) and Convolutional Neural Networks (CNNs) in tomato ripeness classification. Traditional CNNs and SE networks rely on deterministic pooling strategies that can be suboptimal for capturing diverse visual features such as texture and color gradients in fruit. A novel fuzzy pooling system is proposed that dynamically adjusts the fusion ratio of min and max pooling based on real-time texture analysis of feature maps via a Laplacian filter. This mechanism was integrated into Darknet53, ResNet34, and ResNet101 architectures, both directly and within the SE attention block. The method was evaluated on a five-class tomato ripeness dataset using 10-fold cross-validation against traditional and fixed-weight pooling methods. The fuzzy and mixed pooling approach demonstrated superior performance on certain models which include ESEnet: Darknet53 + SE (Fuzzy pool) achieved an average F1 score across folds of 0.8429, outperforming Darknet53 + baseline SE (Average pool) at 0.8350; Extended Net: ResNet34 + Fuzzy pool achieved an average F1 score across folds of 0.8305, outperforming baseline ResNet34 at 0.8130 and ResNet101 + Fuzzy pool achieved an average F1 score across folds of 0.8284, outperforming baseline ResNet101 at 0.8267. This improvement was realized with minimal increase in computational overhead. Specifically, measurements of model parameters (M) and FLOPs (GMACs) showed only a slight increase across all architectures after integrating SEnet into the system, and zero increase after integrating the Laplacian filter and fuzzy logic. The study concludes that logic-based adaptive pooling is a robust strategy that improves feature representation in CNNs, achieving a balance between classification accuracy and computational cost. These findings offer practical value for developing more reliable automated harvesting systems and provide a generalizable enhancement for a wide range of computer vision classification tasks.