YOLOX-Dense-CT: a Detection Algorithm for Cherry Tomatoes Based on YOLOX and DenseNet

The detection of cherry tomatoes has a great significance for robotic harvesting. However, uneven environment conditions, such as branch and leaf occlusion, cluster of fruits, and so on, have made the cherry tomatoes detection very challenging. This paper proposes an effective cherry tomato detection algorithm called YOLOX-Dense-CT to solve these problems. To be specific, the DenseNet network is treated as the basic backbone of the original YOLOX to make the whole network suitable for the cherry tomatoes. Moreover, the convolutional block attention module (CBAM) attention mechanism is applied to make the features from the backbone more fused with the Neck part. As suggested by the experimental results, the proposed YOLOX-Dense-CT model is effective in detecting cherry tomatoes, with the mean average precision (mAP) reaching 94.80%, which is 4.02% higher than the original YOLOX-L model. Meanwhile, the number of parameters is only 34.6 M, which is 19.6 M lower than the YOLOX-L. Furthermore, the YOLOX-Dense-CT is compared to general target detection models and it has the best detection performance. In summary, the proposed method can well meet the requirements of high accuracy detection and provide a strategy for the cherry tomato detection system.