Selective Spatio-Temporal Fruit Inspector (S$^2$FI): A Mamba-based Framework for Rotational Fruit Quality Assessment

Hao Tian; Jie Hu; Guoqi Shan; Fan Wang; Wenzhuo Zhang; Youshan Zhang

cs.CVOct 9, 2026

Selective Spatio-Temporal Fruit Inspector (S $^2$ FI): A Mamba-based Framework for Rotational Fruit Quality Assessment

Authors: Hao Tian, Jie Hu, Guoqi Shan, Fan Wang, Wenzhuo Zhang, Youshan Zhang

Organizations: School of Artificial Intelligence, Chuzhou University, Anhui, China · School of Computer and Information Engineering, Chuzhou University, China

Abstract

The global fruit and vegetable industry confronts severe post-harvest sorting challenges for small and medium-scale growers. Large intelligent sorting systems are too costly, while manual sorting suffers from low efficiency, inconsistent standards, and high labor demand. To bridge this gap between expensive equipment and inefficient manual work, this study develops a low-cost, high-efficiency, lightweight fruit sorting system driven by deep learning, suitable for small and medium growers. We propose the Selective Spatio-Temporal Fruit Inspector (S $^2$ FI) architecture, which combines a 3D convolutional encoder with a Mamba-based backbone.It efficiently extracts spatiotemporal features, adaptively weighs informative frames during fruit rotation, and models long-range dependencies with linear complexity, reducing motion blur and occlusion. The hardware uses a dual-conveyor and curved rail design, supporting household power, fast deployment, and lightweight embedded deployment to improve affordability. We also built a conveyor-based fruit quality video dataset with 6,460 samples. Experiments on the held-out test set show that our method achieves the best trade-off among accuracy, efficiency, and practicality, outperforming the compared baselines on the test set, achieving an excellent balance of accuracy, speed, and practicality for real-world sorting.

Figures & tables

Figure 1: Architecture of the proposed S 2 FI model. From left to right: (a) multi-dimensional feature fusion via color, defect, and shape cues from 3D convolutional encoder; (b) Mamba blocks with auxiliary supervision heads for deep temporal modeling; (c) multi-statistic fusion (avg/max/std) to aggregate discriminative temporal patterns; (d) final classification outputting fruit quality categories.

Category	Training Set	Validation Set	Test Set	Total
Good Fruit	2729	390	779	3898
Substandard Fruit	1441	206	411	2058
Rotten Fruit	353	50	101	504
Total	4523	646	1291	6460

Table 1: Summary of the Fruit Sorting Dataset

Figure 2: Fruit Morphological Classification Standards

Model	Good (%)	Medium (%)	Bad (%)	MCA (%)	Macro F1	Params (M)
R3D [ 24 ]	24.1	75.2	100.0	66.5	0.608	900.7
I3D [ 25 ]	45.4	62.1	100.0	69.2	0.672	529.2
C3D [ 26 ]	100.0	0.0	0.0	33.3	0.251	73.8
Transformer [ 27 ]	100.0	0.0	0.0	33.3	0.251	274.9
MoCLIP-Lite [ 28 ]	0.0	0.0	100.0	33.3	0.2084	28.5
DVFL-Net [ 29 ]	67.2	87.4	100.0	84.9	0.824	32.7

Table 2: Validation performance comparison of fruit quality classification models.Note: Medium corresponds to substandard fruit; Bad corresponds to rotten fruit.Some baseline models suffer from training collapse, tending to predict the majority class and yielding near-zero accuracy on minority categories.

Model	Good (%)	Medium (%)	Bad (%)	MCA (%)	Macro F1	Params (M)
R3D [ 24 ]	66.1	34.9	1.0	67.0	0.589	900.7
I3D [ 25 ]	65.5	35.2	96.8	65.8	0.632	529.2
C3D [ 26 ]	100.0	0.0	0.0	33.3	0.251	73.8
Transformer [ 27 ]	100.0	0.0	0.0	33.3	0.251	274.9
MoCLIP-Lite [ 28 ]	0.07	36.9	14.0	26.3	0.235	28.5
DVFL-Net [ 29 ]	55.5	20.7	43.7	40.6	0.424	32.7

Table 3: Testing performance comparison of fruit quality classification models. Note: Medium corresponds to substandard fruit; Bad corresponds to rotten fruit. Some baseline models suffer from training collapse, tending to predict the majority class and yielding near-zero accuracy on minority categories.

Variant	Good (%)	Medium (%)	Bad (%)	MCA (%)	Macro F1
w/o 3D Encoder	74.2	0.3	99.8	58.1	0.389
w/o Mamba Backbone	98.3	3.9	99.8	67.3	0.587
w/o Multi-stat. Fusion	96.4	12.9	100.0	69.7	0.623
w/o Auxiliary Head	87.9	24.5	100.0	70.8	0.601
w/o SACL	97.7	2.5	93.9	64.7	0.563
Full $S^{2}$ FI	71.3	48.1	100.0	73.2	0.723

Table 4: Ablation study of individual components in

S^{2}

FI.

Explore similar work

FruitEnsemble: MLLM-Guided Arbitration for Heterogeneous ensemble in Fine-Grained Fruit Recognition
May 20, 2026Enhui Yu, Junhui Li, Ruitong Lu +2Agricultural Image Analysis Ensemble Learning
University of Science and Technology Liaoning · Chuzhou University · Yeshiva University
Non-Destructive Prediction of Fruit Ripeness and Firmness Using Hyperspectral Imaging and Lightweight Machine Learning Models
Apr 5, 2026Phongsakon Mark Konrad, Casper Kunstmann-Olsen, Jacek Fiutowski +1Agricultural Image Analysis Hyperspectral Imaging
The Maersk Mc-Kinney Moller Institute, SDU Centre for Industrial Software, University of Southern Denmark, Sønderborg, 6400, Denmark · Mads Clausen Institute, SDU NanoSyd, University of Southern Denmark, Sønderborg, 6400, Denmark
Towards Practical Precision Agriculture: Real-Time Fruit Detection and Video Analytics on Embedded Edge Hardware
Sep 11, 2026Ivica Dimitrovski, Vlatko Spasev, Ivan Kitanovski +2Agricultural Image Analysis Real-Time Object Detection
Faculty of Computer Science and Engineering, University Ss Cyril and Methodius, ul. Rudzer Boshkovikj 16, P.O. 393, Skopje, 100, North Macedonia · Faculty of Agricultural Sciences and Food, University Ss Cyril and Methodius, 16-ta Makedonska brigada br. 3, Skopje, 100, North Macedonia

May 20, 2026J/K move · Enter open · S save

FruitEnsemble: MLLM-Guided Arbitration for Heterogeneous ensemble in Fine-Grained Fruit Recognition

Enhui Yu, Junhui Li, Ruitong Lu, Jialu Li+1

University of Science and Technology Liaoning · Chuzhou University · Yeshiva University

Agricultural Image Analysis Ensemble Learning Fine-Grained Image Classification Fine-Grained Visual Perception Image Classification

Open paper AskPDF arXiv

Fine-grained fruit classification is a critical yet challenging task in agricultural computer vision, primarily hindered by a severe shortage of high-quality datasets and the high visual similarity between classes. To address these challenges, we first constructed a comprehensive dataset comprising 306 fruit categories with 116,233 samples. Moreover, we propose FruitEnsemble, a practical two-stage dynamic inference framework designed to overcome the generalization limitations of static single-model architectures. In the first stage, FruitEnsemble employs a validation-calibrated weighted ensemble of heterogeneous backbones to generate a robust Top-3 candidate pool. To tackle difficult samples, we introduce an expert arbitration mechanism: when ensemble confidence falls below 0.6, a multimodal large language model (MLLM) is triggered to perform rigorous visual verification by integrating external botanical descriptions using Chain-of-Thought (CoT) reasoning. Furthermore, we optimized the training pipeline with a hard sample-aware joint loss. Extensive experiments demonstrate that FruitEnsemble achieves a classification accuracy of 70.49% and outperforms existing state-of-the-art models. Our framework provides an efficient, deployment-oriented solution for real-world agricultural visual sorting and quality inspection tasks.

Abstract

Explore similar work

FruitEnsemble: MLLM-Guided Arbitration for Heterogeneous ensemble in Fine-Grained Fruit Recognition

Non-Destructive Prediction of Fruit Ripeness and Firmness Using Hyperspectral Imaging and Lightweight Machine Learning Models

Towards Practical Precision Agriculture: Real-Time Fruit Detection and Video Analytics on Embedded Edge Hardware