This paper presents an improved tomato detection algorithm, SBCS-YOLOv5s, designed to enhance the accuracy and efficiency of tomato detection in natural environments. The algorithm integrates several advanced modules into the YOLOv5s model to address challenges such as uneven illumination, occlusions, and overlapping fruits. Key modifications include: 1. **C3SE Module**: Combining the SE attention module with the C3 module to improve feature extraction. 2. **BiFPN**: Replacing the PANet with a weighted Bi-directional Feature Pyramid Network to enhance feature fusion. 3. **CARAFE Module**: Replacing the regular up-sampling operator with CARAFE to improve spatial detail and semantic information. 4. **Soft-NMS Algorithm**: Replacing the conventional NMS algorithm with Soft-NMS to better handle overlapping and occluded fruits. The experimental results show that SBCS-YOLOv5s achieved a mean average precision (mAP) of 87.7%, a 3.5% improvement over the original YOLOv5s model. The detection speed was 2.6 ms per image, outperforming other state-of-the-art detection algorithms. The model demonstrated robust performance under various conditions, including different lighting and occlusion levels. The study concludes by highlighting the potential of SBCS-YOLOv5s for advancing tomato harvesting robots and future work on incorporating contextual and maturity information.This paper presents an improved tomato detection algorithm, SBCS-YOLOv5s, designed to enhance the accuracy and efficiency of tomato detection in natural environments. The algorithm integrates several advanced modules into the YOLOv5s model to address challenges such as uneven illumination, occlusions, and overlapping fruits. Key modifications include: 1. **C3SE Module**: Combining the SE attention module with the C3 module to improve feature extraction. 2. **BiFPN**: Replacing the PANet with a weighted Bi-directional Feature Pyramid Network to enhance feature fusion. 3. **CARAFE Module**: Replacing the regular up-sampling operator with CARAFE to improve spatial detail and semantic information. 4. **Soft-NMS Algorithm**: Replacing the conventional NMS algorithm with Soft-NMS to better handle overlapping and occluded fruits. The experimental results show that SBCS-YOLOv5s achieved a mean average precision (mAP) of 87.7%, a 3.5% improvement over the original YOLOv5s model. The detection speed was 2.6 ms per image, outperforming other state-of-the-art detection algorithms. The model demonstrated robust performance under various conditions, including different lighting and occlusion levels. The study concludes by highlighting the potential of SBCS-YOLOv5s for advancing tomato harvesting robots and future work on incorporating contextual and maturity information.