Abstract: Abstract: Fruit recognition and segmentation using deep neural networks have widely been contributed to the operation of picking robots in modern agriculture. However, the most current models present a low accuracy of recognition with a low running speed, due mainly to a large number of network parameters and calculations. In this study, a high-resolution segmentation was proposed for the different ripeness of tomatoes under a greenhouse environment using improved Mask R-CNN. Firstly, a Cross Stage Partial Network (CSPNet) was used to merge with Residual Network (ResNet) in the Mask R-CNN model. Cross-stage splitting and cascading strategies were contributed to reducing the repeated features in the backpropagation process for a higher accuracy rate, while reducing the number of network calculations. Secondly, the cross-entropy loss function with weight factor was utilized to calculate the mask loss for the better segmentation effect of the model, due to the imbalance of the whole sample. An experiment was also performed on the test sets of tomato fruits with three ripeness levels. The results showed that the improved Mask R-CNN model with CSP-ResNet50 as the backbone network presented the mean average precision of 95.45%, the precision of 95.25%, the recall of 87.43%, F1-score of 0.912, and average segmentation time was 0.658 s. Furthermore, the mean average precision increased by 16.44, 14.95, and 2.29 percentage points, respectively, compared with the Pyramid Scene Parsing Network (PSPNet), DeepLab v3+, and Mask R-CNN with ResNet50 as the backbone network. Nevertheless, the average segmentation time increased by 14.83% and 27.52%, respectively, compared with PSPNet and DeepLab v3+. More importantly, the average segmentation time of improved Mask R-CNN with CSP-ResNet50 as the backbone network was reduced by 1.98%, compared with Mask R-CNN with ResNet50 as the backbone network. Additionally, the new model performed well in the segmentation of green and half-ripe tomato fruits under different light intensities, especially under low light, compared with PSPNet and DeepLab v3+. Finally, the improved Mask R-CNN model with CSP-ResNet50 as the backbone network was deployed to the picking robot, in order to verify the recognition and segmentation effect on different ripeness of tomato fruits in large glass greenhouses. In a low overlap rate of tomato fruits, the model identified the number of tomato fruits consistent with manual detection, where the accuracy was more than 90%. When the occlusion or overlap rate of tomato fruits exceeded 70%, particularly when the target was far away, the accuracy of 66.67% was achieved in the improved Mask R-CNN model, indicating a large gap with manual detection. Only a few features with the blur pixels were attributed to the difficulty to extract the shape and color features of tomato fruits. In addition, low light also posed a great challenge on recognition difficulty. Correspondingly, it was more difficult to pick tomatoes for the picking robot, particularly a relatively low success rate of picking, as the overlap was more serious. Fortunately, the picking success rate improved greatly, as the occlusions reduced. Consequently, the integrated multiple technologies (such as image acquisition equipment, the performance of the model, the execution end design of robotic arm, and automatic mechanization) can widely be expected to effectively improve the picking rate of mature tomatoes under the complex environment of a specific greenhouse. The new model also demonstrated strong robustness and applicability for the precise operation of tomato-picking robots in various complex environments.