💻 computer science

MobileNetV2-KAN: A Lightweight Nonlinear Network for Cross- Dataset Fruit Maturity Recognition

This paper introduces MobileNetV2-KAN+, a lightweight 3.65-million-parameter architecture that combines a MobileNetV2 backbone with a bounded Kolmogorov-Arnold Network classifier to achieve robust, well-calibrated cross-dataset fruit maturity recognition without relying on pretrained weights or large model sizes.

Original authors: Zhengnan Caiᵃ, Yuelong Chenᵃ, YaXuan Huangᵃ, Ziqi Fanᵃ, Xiaodong Wangᵃ

Published 2026-10-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhengnan Caiᵃ, Yuelong Chenᵃ, YaXuan Huangᵃ, Ziqi Fanᵃ, Xiaodong Wangᵃ

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quiet rhythm of an orchard, the difference between a harvest that feeds a family and one that goes to waste often comes down to a single, subtle shift in color. For centuries, farmers have relied on experience to judge when a fruit is ready, but as agriculture moves toward automation, robots and sensors must learn to see what the human eye sees. This is the challenge of computer vision in farming: teaching machines to distinguish a ripe lychee from an unripe one, or a yellow banana from a spotted, overripe one, amidst the chaotic reality of wind-blown leaves, shifting shadows, and uneven lighting. While modern computers have become incredibly good at recognizing objects in controlled, laboratory settings, they often stumble when faced with the messy, unpredictable conditions of a real field. The gap between a model that works perfectly in a lab and one that works reliably in an orchard remains a significant hurdle for smart agriculture.

Researchers at Xiamen University Tan Kah Kee College have proposed a new approach to bridge this gap, focusing on a specific type of artificial intelligence architecture designed to be both lightweight and highly adaptable. Their work, centered on a model they call MobileNetV2-KAN, seeks to solve the problem of "cross-dataset" recognition. In simple terms, this means building a system that can learn from one set of images—perhaps taken with one camera under specific lighting—and still perform well when shown images taken with a different camera, in a different location, or under different weather conditions. The team argues that many current systems are too rigid; they memorize the specific conditions of their training data rather than learning the fundamental visual traits of fruit maturity. To fix this, they combined a proven, efficient image-processing backbone with a newer, more flexible mathematical structure known as a Kolmogorov-Arnold Network, or KAN, which excels at learning complex, non-linear patterns without needing massive amounts of computing power.

The core of their experiment involved teaching the model to recognize the maturity of lychees, a fruit that transitions from green to red as it ripens. The researchers did not simply feed the computer thousands of pictures and hope for the best. Instead, they constructed a rigorous test to see if the model could truly generalize. They trained the system on a large collection of lychee images gathered from one source, but then tested it on a completely separate set of images taken from a different dataset, captured with different equipment and under different conditions. This "cross-dataset" test is crucial because it mimics the real-world scenario where a farmer might use a model trained on data from one region to manage an orchard in another. The results were revealing. While many standard models performed well when tested on data similar to what they were trained on, their accuracy dropped significantly when faced with the new, unseen conditions. The MobileNetV2-KAN model, however, maintained a much higher level of stability. When tested on the independent lychee dataset, it achieved an accuracy of roughly 77.57%, outperforming several established, heavier models that struggled to adapt to the new environment.

What makes this result particularly notable is the efficiency of the solution. The researchers did not simply throw more computing power at the problem. The model they built is "lightweight," meaning it has a relatively small number of parameters, or internal settings, compared to the massive networks often used in deep learning. This is essential for deployment in the field, where devices like drones or handheld scanners have limited battery life and processing capacity. By replacing the standard final layer of the network with the KAN structure, the model gained a better ability to map the subtle, non-linear differences between fruit stages. Instead of relying on rigid, pre-defined rules, the KAN component learns flexible, one-dimensional functions that can adjust to the nuances of the data. This allowed the model to capture the fine-grained details of fruit appearance—such as the specific shade of red on a ripe lychee—without becoming confused by the background noise of leaves or shadows.

To ensure their findings were not a fluke, the team also tested the model on a completely different type of fruit: bananas. This dataset presented a new challenge, as bananas change from green to yellow and then develop brown spots as they overripen, a visual progression distinct from the red-shift of lychees. The researchers applied the exact same model and training strategy without making any special adjustments for the new fruit. The model adapted successfully, achieving a test accuracy of roughly 94.91% on the banana dataset. This cross-crop success suggests that the method is not limited to recognizing one specific fruit, but is a robust approach that can be applied to various agricultural tasks. The model demonstrated that it could learn the underlying principles of fruit maturity recognition rather than just memorizing the specific pixels of the training images.

The study also explored how the amount of training data affected the model's performance. In many machine learning tasks, the assumption is that more data always leads to better results. However, in this specific context, the researchers found that the model remained stable even when trained on smaller subsets of data. When they reduced the amount of training data to just 20% of the available set, the model still performed competitively, whereas other models showed more volatility. This suggests that the KAN structure is particularly effective at learning from limited samples, a critical advantage in agriculture where gathering large, perfectly labeled datasets can be difficult and expensive. The researchers noted that while the model is powerful, its success depends heavily on the quality of the data; if the images are too noisy or the labels are inconsistent, the model's ability to generalize can suffer.

Ultimately, this research points toward a future where agricultural AI is not just powerful, but practical and portable. By proving that a lightweight, hybrid model can handle the unpredictability of real-world environments better than traditional, heavier networks, the authors offer a viable path forward for automated harvesting and crop management. The work suggests that the key to solving the "lab-to-field" gap is not necessarily building bigger, more complex systems, but rather designing smarter, more flexible ones that can learn the true essence of a fruit's maturity, regardless of where or how the image was taken. As the technology matures, such systems could allow farmers to deploy intelligent tools that work reliably across different seasons, locations, and crop types, turning the complex art of fruit recognition into a precise, automated science.

Technical Summary: MobileNetV2-KAN for Cross-Dataset Fruit Maturity Recognition

1. Problem Statement

Automatic crop recognition is vital for smart agriculture, yet existing deep learning models face two primary bottlenecks in real-world deployment:

Deployment Efficiency: Many state-of-the-art models rely on large parameter counts and complex structures, making them inefficient for resource-constrained edge devices in field environments.
Generalization Gap ("Lab-Field" Discrepancy): Current evaluation paradigms predominantly rely on same-distribution partitioning (training and testing on data from identical acquisition conditions). This leads to performance saturation that masks a model's true adaptability. Consequently, models often suffer significant performance degradation when applied to cross-dataset scenarios involving different lighting, backgrounds, imaging equipment, or crop varieties.

The paper specifically addresses the challenge of recognizing fruit maturity (e.g., lychee and banana) under these distribution shifts while maintaining a lightweight architecture suitable for edge inference.

2. Methodology

The authors propose MobileNetV2-KAN, a hybrid architecture designed to balance computational efficiency with enhanced nonlinear modeling capabilities.

2.1 Architecture Overview

The model consists of two main components:

Feature Extraction Backbone: A standard MobileNetV2 network is utilized to extract spatial features. It employs inverted residual blocks with depthwise separable convolutions to minimize computational overhead. The input images are resized to 64×64 pixels, and the backbone outputs a 1280-dimensional feature map which is flattened into a 5120-dimensional vector.
Classification Head (KAN): Instead of traditional fully connected (MLP) layers, the classification head integrates Kolmogorov-Arnold Networks (KAN).
- Mechanism: Based on the Kolmogorov-Arnold theorem, the model replaces fixed weights with learnable one-dimensional functions. Specifically, the authors implement KANLinear layers using B-spline basis functions (cubic, degree $p=3$ ).
- Function: For an input vector $x$ , the output is computed as a sum of learnable 1D functions $\phi_{j,i}(x_i)$ rather than a simple linear combination. This allows the network to learn smooth, nonlinear mappings directly from the data, enhancing its ability to capture fine-grained maturity features without a massive increase in parameters.

2.2 Experimental Design & Data

Datasets:
- Lychee Maturity Dataset: Derived from Zhao et al. (2025), featuring orchard images with YOLO-format annotations. It is split into Training, Validation, Test, and an independent Test1 (cross-dataset) set derived from a different source (An et al., 2025) to evaluate distribution shift robustness.
- Banana Maturity Dataset: Sourced from the Alibaba Tianchi platform, used to verify cross-crop generalization.
Training Strategy:
- Preprocessing: Instance-level cropping is performed. No forced resizing is applied during dataset construction to preserve spatial characteristics, though input images are resized to 64×64 during the loading phase.
- Augmentation: Minimal augmentation (random rotation, horizontal flip, slight color jitter) is applied only during training to avoid artificial bias.
- Optimization: The model uses Stochastic Gradient Descent (SGD) with momentum, weight decay, and a cosine annealing learning rate schedule. Label Smoothing (LS) is employed to mitigate overfitting in small-sample scenarios.
- Evaluation: Experiments are conducted with multiple random seeds to report mean ± standard deviation, ensuring reproducibility.

3. Key Contributions

Novel Hybrid Architecture: The proposal of MobileNetV2-KAN, which integrates a learnable nonlinear mapping module (KAN) into a lightweight convolutional backbone. This aims to enhance fine-grained decision-making for maturity classification without the computational cost of heavy transformers or large CNNs.
Rigorous Cross-Dataset Evaluation: The study introduces a strict evaluation protocol using an independent cross-source test set (Test1) for lychee and a separate crop dataset (banana). This moves beyond standard same-distribution validation to explicitly measure "lab-field" generalization.
Stability Analysis under Data Scarcity: The authors investigate model performance across varying training subset proportions (0.2, 0.4, 0.6) and random seeds, demonstrating that the proposed method maintains stability even with limited training data.

4. Experimental Results

Cross-Dataset Performance (Lychee):
- On the independent cross-dataset test set (Test1), MobileNetV2-KAN achieved a mean accuracy of 77.57% (with a training subset ratio of 0.2).
- This outperformed baseline models (MobileNetV2, ResNet50, VGG16, EfficientNetB3), which generally showed lower accuracy and higher variance (e.g., MobileNetV2 achieved 67.55% under the same conditions).
- The proposed model demonstrated smaller performance fluctuations across different training subset ratios, indicating superior robustness to distribution shifts.
Same-Distribution Performance:
- On the same-distribution test set, all models achieved high accuracy (>98%), confirming that the performance gap is primarily driven by generalization challenges rather than basic classification capability.
Cross-Crop Generalization (Banana):
- When applied to the banana dataset without architecture modification, MobileNetV2-KAN achieved a test accuracy of 94.91% (ratio 0.4), comparable to larger models like ResNet50 and EfficientNetB3, validating its transferability across different fruit types.
Efficiency:
- The model maintains a parameter count of approximately 6.16 million (23.51 MB), which is larger than standard MobileNetV2 (2.2M) but significantly smaller than ResNet50 (23.5M) and VGG16 (134M), offering a trade-off between efficiency and enhanced nonlinear expressiveness.

5. Significance and Claims

The paper claims that MobileNetV2-KAN offers a feasible solution for intelligent recognition systems in practical agricultural scenarios where:

Data Diversity is Limited: The model's ability to perform well with smaller training subsets (ratio 0.2) suggests that the KAN head's nonlinear approximation capabilities can compensate for data scarcity.
Environmental Conditions Vary: The superior performance on the cross-dataset test set indicates that the learnable 1D functions in the KAN head are better at modeling the complex, non-linear decision boundaries required to handle lighting and background variations than standard linear layers.
Resource Constraints Exist: By combining a lightweight backbone with an efficient nonlinear head, the model achieves a balance between computational overhead and generalization robustness.

The authors conclude that while same-distribution tests verify training stability, cross-dataset and cross-crop evaluations are essential to truly assess a model's readiness for real-world agricultural deployment. The study highlights that the KAN module's flexibility in function approximation is a key factor in improving generalization without requiring massive parameter increases.

6. Limitations and Future Work

The authors modestly note that the current study is limited to binary classification tasks (ripe vs. unripe) and does not address multi-class maturity stages. Future work aims to:

Extend the B-spline degree adaptability mechanism for more complex multi-class and cross-species tasks.
Introduce more advanced data augmentation and cross-crop training strategies to further enhance robustness.
Explore the model's performance in multi-class scenarios beyond binary maturity detection.