基于多尺度融合与动态特征交互的仓储环境马铃薯缺陷检测(MDS-DETR)

Logical Operator Operator

Search Text

Search Type

add_circle_outline

remove_circle_outline

Article Menu

Font Type:

Arial Georgia Verdana

Font Size:

Aa Aa Aa

Line Spacing:

  

Column Width:

  

Background:

Open Access Article

by
Danyang Lv

1,2,
Ang Zhao

2,3,
Shuo Han

2,3,
Ranbing Yang

2,3,*,
Guohai Zhang

4 and
Xiaohui Yang

4

1

School of Information and Communication Engineering, Hainan University, Haikou 570228, China

2

Key Laboratory of Tropical Intelligent Agricultural Equipment, Hainan University, Ministry of Agriculture and Rural Affairs, Danzhou 571737, China

3

School of Mechanical and Electrical Engineering, Hainan University, Haikou 570228, China

4

School of Agricultural Engineering and Food Science, Shandong University of Technology, Zibo 255049, China

Author to whom correspondence should be addressed.

Submission received: 15 July 2026 / Revised: 12 August 2026 / Accepted: 12 August 2026 / Published: 13 August 2026

Abstract

Aiming at the challenges of potato storage scenarios, severe target stacking and occlusion, and large variations in defect characteristics in potato storage environments, a potato defect detection method named MDS-DETR was proposed, and its deployment on a robotic sorting platform was validated. First, potato images under different illumination conditions and stacking states were collected in potato storage warehouses to construct a potato defect dataset containing defects such as black spot, sprouting, dry rot, decay, green skin, and cracking. Subsequently, according to the characteristics of object detection tasks in storage scenarios, an MCF module was designed to strengthen the extraction and integration of contextual features across different spatial scales. An AIFI-DyMona structure, namely Anchor-free Instance Feature Interaction with Dynamic Mona, was constructed to improve the stability of feature representation under different illumination conditions. In addition, the Shape-IoU regression strategy was incorporated to improve the network sensitivity to irregular defect contours and geometric characteristics. Validation experiments demonstrated that the proposed MDS-DETR framework achieved 96.3% [email protected], with only 14.2 M parameters and 42.9 G FLOPs. Compared with several representative object detection algorithms, the proposed method exhibited superior recognition capability and stronger robustness under complicated storage conditions, while also showing better suppression of missed and incorrect detections. To further evaluate its practical applicability, the trained network was integrated into an intelligent potato sorting robot and tested in real warehouse scenarios. Experimental observations indicated that the robotic system consistently maintained a sorting accuracy exceeding 95%, demonstrating the effectiveness and practical deployment potential of the proposed approach for potato storage applications. This study can provide a reference for intelligent detection and automated sorting in potato storage processes.

1. Introduction

Potato is recognized as an essential crop in global agriculture due to its high productivity, wide environmental adaptability, and abundant nutritional content [1]. Driven by advances in large-scale cultivation methods and centralized storage systems, potatoes are increasingly preserved in cold storage facilities or ventilated warehouses after harvest for medium to long-term storage, thereby supporting off-season marketing and maintaining a stable market supply. However, during the storage process, fluctuations in temperature, humidity, and ventilation conditions can easily lead to quality problems such as decay and sprouting [2]. The occurrence of surface damage and pathological symptoms can significantly decrease the marketability of potatoes while increasing the risk of cross-infection during storage, which further threatens the overall quality and safety of stored products [3]. Therefore, efficient and accurate quality detection of potatoes during the storage stage is of great importance for reducing storage losses and improving economic benefits.

Traditional crop quality inspection mainly relies on manual judgment, where operators identify defects such as decay and sprouting through visual observation [4]. Although this approach offers a certain degree of flexibility, its reliance on manual inspection leads to considerable subjectivity and labor consumption. To improve detection efficiency, some studies have adopted traditional image processing techniques to identify crop appearance characteristics through the analysis of visual attributes, including chromatic, textural, and contour information. To overcome the limitations of conventional manual grading in citrus inspection, Gong et al. [5] developed an automated defect recognition approach using machine vision technology, which significantly enhanced sorting speed and detection reliability. The method achieved defect extraction through Hue, Saturation, and Value (HSV) color space transformation, edge detection, and morphological processing, obtaining an overall recognition accuracy of 92%. Li et al. [6] presented an automated detection approach for pineapple, which combined color and depth information with multiple image processing techniques to segment damaged regions and quantitatively calculate the damaged area through binarization processing. Nevertheless, these approaches are highly susceptible to lighting conditions and environmental noise, which limits their stability in complicated storage scenes and reduces their ability to accurately detect irregular defects, including rot and sprouting in potatoes.

Driven by the rapid advancement of smart agriculture and digital farming technologies, neural network-based visual recognition methods have been applied in agricultural product inspection and quality evaluation [7]. Models represented by convolutional neural networks and Transformers can automatically learn multi-level feature representations and have demonstrated significant advantages in quality assessment applications [8,9]. In recent years, intelligent visual recognition techniques have been increasingly introduced into potato inspection tasks. For instance, Wang et al. [10] established a dedicated potato image dataset and conducted comparative experiments on three convolutional neural network architectures, demonstrating the feasibility of deep learning for identifying potato defects. To improve the accuracy of potato appearance inspection, Liao et al. [11] designed an external quality detection model named HCRP-YOLO, which achieved 95.8% mean average precision in experimental evaluations. However, most existing studies focus on ideal illumination conditions and controllable indoor scenarios, while research on applications in storage environments remains relatively limited.

In real storage scenarios, potatoes are usually stacked or scattered, resulting in mutual occlusion among individuals. In addition, soil may adhere to the potato surface and color changes may occur, leading to blurred object boundaries and making stable segmentation difficult for traditional methods. Meanwhile, illumination conditions in storage warehouses are often uneven or weak, which can easily produce shadows or locally dark regions, thereby affecting image quality. Furthermore, as storage time increases, the appearance characteristics of potatoes dynamically change, such as gradual sprout growth and expansion of decayed areas. Therefore, achieving high-precision identification of decayed and sprouted potatoes under complex storage environments has become a critical issue in potato storage management.

This study focuses on potato defect detection tasks in storage scenarios. The primary work and innovations of this study can be summarized as follows:

(1) A potato image dataset close to real storage environments was constructed, covering different illumination conditions, stacking states, and defect types.

(2) The Real-Time Detection Transformer (RT-DETR) framework was further refined to strengthen its ability to capture intricate feature information, leading to more reliable recognition performance under weak illumination and severe occlusion conditions.

(3) The developed detection algorithm was integrated into an intelligent potato sorting robot, and practical experiments in warehouse environments were carried out to assess its real-world identification and sorting capability.

2. Materials and Methods

2.1. Research Framework

Figure 1 illustrates the workflow of this study, which mainly includes potato image collection and dataset establishment, network optimization and training, comparative evaluation of different detection models, and practical verification through robotic sorting experiments.

Figure 1. Overall framework.

2.2. Potato Image Collection and Dataset Preparation

The potato images were collected in March 2026 from a controlled atmosphere potato storage warehouse located in Guyuan County, Zhangjiakou City, Hebei Province, China, as shown in Figure 2. The stored potato varieties mainly included Xisen 3 and Xisen 6. Image acquisition was performed using a BNC100 industrial camera (Beijing Xinhong Teaching Equipment Co., Ltd., Beijing, China). The BNC100 industrial camera is equipped with a 1/2.7-inch CMOS sensor, providing an image resolution of 1600 × 1200 pixels.

Figure 2. Data acquisition site.

To improve the adaptability of the dataset to real storage environments, complex warehouse scene characteristics were fully considered during the image acquisition process. The collected samples included both stacked and scattered potatoes. In addition, the dataset included both normal and low-light conditions, allowing the model to achieve more stable performance across diverse lighting environments. The collected defect types mainly included black spot, sprouting, dry rot, rot, greening, and cracking. Additionally, some images were obtained from the Internet.

After image acquisition, samples with severe blur, heavy occlusion, or incomplete information were manually discarded during the data screening process, and the remaining images were annotated in LabelImg for subsequent model training. Since the primary objective of this study was defect detection for stored potatoes, multiple defect types, including black spot, sprouting, dry rot, rot, greening, and cracking, were consolidated into a unified category labeled “Defect” to alleviate the impact of category variability and enhance the model’s capability to identify defective potatoes under complex storage conditions. Normal potatoes were not assigned a separate category and were not annotated.

Finally, a potato defect dataset containing 3540 images was constructed. The dataset construction process is shown in Figure 3.

Figure 3. Dataset construction process.

2.3. MDS-DETR Detection Method

Proposed by the Baidu research team [12], RT-DETR is a Transformer-based real-time detection framework that achieves faster inference and higher detection efficiency than conventional DETR architectures while preserving excellent detection accuracy. Although RT-DETR has clear advantages in end-to-end detection frameworks and inference efficiency, it still has certain limitations in storage environments. Specifically, the original backbone network has insufficient representation capability for fine-grained textures and small defects. In addition, the feature interaction process is sensitive to illumination variations, and the feature perception mechanism cannot adequately describe the geometric shape information of targets. These factors may lead to decreased recognition accuracy in potato defect detection tasks.

The proposed method optimizes the network architecture through targeted optimization of feature representation, information interaction, and feature perception to enhance overall detection performance. Based on these improvements, an enhanced detection model named MDS-DETR is constructed, as shown in Figure 4.

Figure 4. Overall structure of MDS-DETR.

2.3.1. MCF

In the baseline RT-DETR model, the BasicBlock is mainly responsible for feature extraction. However, this structure primarily depends on local convolution operations, and its effective receptive field is relatively limited. As a result, the network has weakened its discrimination capability when dealing with potatoes with similar appearances or complex backgrounds.

As shown in Figure 5, this study introduces a Multi-scale Context Fusion module named MCF into the backbone network as a substitute for the conventional BasicBlock structure. The purpose of this module is to strengthen the extraction and integration of contextual information at different scales. On the basis of the residual learning framework, MCF incorporates the Context Guided Block (GCBlock) [13], so that local detail information and broader semantic context can be jointly modeled. GCBlock is composed of four functional components, including a local feature extractor f loc(), a surrounding context extractor f sur(), a joint feature extractor f joi(), and a global context extractor f glo(). The module mainly includes a feature learning stage and a feature refinement stage.

Figure 5. Structure of MCF: (a) MCF; (b) GCBlock.

In the feature learning stage, f loc() focuses on capturing fine-grained spatial details from neighboring regions, whereas f sur() is responsible for modeling broader contextual dependencies and background semantic information around the target area. Specifically, f loc() employs a convolution to capture local spatial features from neighboring regions, while f sur() uses a channel-wise dilated convolution to establish correlations between potato targets and their surrounding background. After that, the two branches are integrated by f joi(*). The calculation process is expressed as follows.

$$
f_{j o i n t} = f_{j o i} \left(f_{l o c} , f_{s u r}\right) = PReLU \left(BN \left[f_{l o c} ; f_{s u r}\right]\right)
$$

(1)

In the feature recalibration stage, f glo(*) is introduced by drawing inspiration from SENet. Adaptive feature weighting is achieved through a channel attention mechanism. First, the channel descriptor is generated:

$$
z = G A P \left(f_{j o i n t}\right) \in R^{B \times C}
$$

(2)

Then, the channel attention weights are calculated:

$$
s = \sigma \left(W_{2} \delta \left(W_{1} z\right)\right)
$$

(3)

where σ(·) denotes the Sigmoid function, δ(·) denotes the ReLU activation function, W 1∈R C/r×C and W 2∈R C×C/r represent the weight matrices, and r denotes the channel reduction ratio.

The generated attention coefficients are subsequently applied to the original feature map through channel-wise weighting, allowing important feature responses to be selectively enhanced:

$$
f_{refined} = f_{joint} \cdot s
$$

(4)

Through this mechanism, feature channels containing critical information were adaptively emphasized, whereas irrelevant responses were weakened, enabling more effective feature optimization under global contextual guidance.

By incorporating the above design, the proposed MCF module strengthens the extraction and integration of contextual features at different scales. Compared with the traditional BasicBlock, MCF simultaneously considers local details, contextual dependencies, and global semantic representations, overcoming the limitation of insufficient contextual modeling in the original backbone network.

2.3.2. AIFI-DyMona

Within the RT-DETR architecture, the normalization operation in the original AIFI module was fixed as Layer Normalization (LayerNorm), which cannot adaptively adjust according to dynamic changes in input features. As a result, the stability of the model under complex illumination conditions is insufficient.

To overcome the above limitation, a Dynamic Mona structure, referred to as DyMona, is introduced in this study and integrated into the Anchor-free Instance Feature Interaction module as a substitute for the original LayerNorm operation. The proposed module integrates the multi-scale feature extraction concept of Mona [14] and introduces the DynamicTanh (DyT) [15] structure to achieve dynamic normalization and extreme value suppression. Based on this improvement, the enhanced AIFI module is named AIFI-DyMona.

DyMona was designed to enhance the fine-grained modeling capability of local features. A normalization and fusion structure was introduced at the top of DyMona. As shown in Figure 6, S 1 and S 2 were employed to adjust the input proportions. In the primary feature extraction branch, depth-wise convolution operations with different kernel sizes were employed to capture spatial information under multiple receptive field ranges. The obtained multi-scale features were subsequently integrated through average fusion to achieve effective interaction and aggregation of cross-scale information. To strengthen channel interactions, a skip connection mechanism was designed in the multi-filter convolution stage. In the first stage, spatial information was strengthened by applying adaptive depth-wise convolution operations, whereas the subsequent phase utilizes dynamically weighted point-wise convolutions to reinforce the network's ability to distinguish informative channel features.

Figure 6. Structure of AIFI-DyMona.

The two convolution operations can be expressed as follows:

$$
f_{d w} = x + a v g \left(\sum_{i = 1}^{3} \omega_{d w}^{i} \hat{\bigotimes} x\right)
$$

(5)

$$
f_{p w} = x + \omega_{p w} \bar{\bigotimes} x
$$

(6)

where x represents the input feature map, ω dw i denotes the convolution weight of the i-th depth-wise convolution branch, ωdw denotes the learnable weight of the point-wise convolution, and

$\hat{\bigotimes}$
and

$\bigotimes$
represent depth-wise convolution and point-wise convolution, respectively. Afterward, nonlinear feature representation was strengthened using the GeLU activation function, followed by an up-projection process to recover the original channel dimensionality. The overall computational process of DyMona can be expressed as follo