An improved Faster R-CNN model for multi-object tomato maturity detection in complex scenarios

Accurate detection of tomato maturity is significant in automatic tomato picking. Although there are many detection methods, they are often sensitive to occlusion, overlap, uneven illumination, and other complex fac-tors, leading to a limited performance in complex environment scenes. To address this problem, this study designed an improved Faster R-CNN model named MatDet for tomato maturity detection. First, MatDet used ResNet-50 as the backbone to improve the representation ability and robustness of the model. Second, RoIAlign was used to obtain more precise bounding boxes in the feature mapping stage. Third, a Path Aggregation Network (PANet) was introduced to address the difficulty of detecting tomato maturity in complex scenarios. Experimental results showed that the proposed model achieved the best detection results in terms of branch occlusion, fruit overlapping and illumination influence under complex scenarios. Specifically, the mean average precision (mAP) of the proposed algorithm is 96.14%, which is better than that of common object detection models. Through many multi-angle comparative experiments, it was confirmed that our method can overcome complex factors such as branch occlusion, fruit overlap and light influence, and achieve the best detection effect. Meanwhile, this research has certain theoretical and practical significance for the intelligent and precise picking of tomatoes, thereby promoting the directional cultivation of crops such as fruits and vegetables, and providing technical support for the development of ecological monitoring technology and ecological planting.