Key points are not available for this paper at this time.
In response to the challenges of drastic illumination variations, large differences in fruit scale, and severe occlusion in real-field environments, this paper proposes a lightweight end-to-end detection model, termed EMBS-DETR, for tomato maturity detection. The proposed method is built upon the RT-DETR-R18 baseline framework, retaining the advantages of global modeling and end-to-end detection enabled by the Transformer architecture, while introducing targeted improvements in feature extraction and multi-scale feature fusion. In the feature extraction stage, a C2f-FDConv module is incorporated to enhance the modeling capability of high-frequency fine-grained features, such as the surface texture and color gradients of tomatoes, while reducing redundant parameter overhead. For high-level semantic representation, an improved parameter-free attention mechanism, SimAM-TF, is designed. By jointly modeling neuron energy functions and color-aware modulation, it effectively enhances feature representation under complex lighting and occlusion conditions. For multi-scale feature fusion, a novel EMBS-FPN structure is proposed. Based on bidirectional feature flow and a multi-scale weighted fusion mechanism, this structure integrates multi-branch receptive field modeling with an efficient upsampling strategy, enabling adaptive fusion of P3–P5 feature layers. This design significantly improves representation stability for objects of varying scales while maintaining model lightweight characteristics. To evaluate the proposed method, a real-field tomato maturity dataset was constructed, consisting of 2327 images collected from facility-grown pink large-fruit tomato varieties widely cultivated in North China. According to agricultural industry standards and physicochemical properties, the dataset is categorized into three classes: immature (796 images), turning stage (718 images), and mature (813 images). Experiments were conducted on an Ubuntu 20.04 platform with an NVIDIA GeForce RTX 3080 Ti GPU. The input resolution was set to 640 × 640. Standard evaluation metrics, including Precision, Recall, mAP@0.5, mAP@0.5:0.95, as well as Params, GFLOPs, and Model Size, were used for comprehensive assessment. The experimental results demonstrate that EMBS-DETR achieves 90.9% Precision, 85.7% Recall, 89.9% mAP@0.5, and 79.8% mAP@0.5:0.95. Meanwhile, with only 37.03 M parameters, 25.2 GFLOPs computational cost, and a model size of 46.3 MB, the proposed model maintains low computational and storage overhead, achieving a favorable balance between accuracy and efficiency. Compared with mainstream YOLO-based models, the proposed method demonstrates superior overall performance in complex field environments, providing effective technical support for automated tomato maturity perception and intelligent visual understanding in precision agriculture.
Yan et al. (Tue,) studied this question.