Maturity identification and category determination method of broccoli based on semantic segmentation models

The critical technology of the broccoli selective harvesting robot centres around the maturity identification and determination of broccoli heads suitable for harvesting. To address this technical challenge, a machine vision method based on semantic segmentation models is proposed in this research. This method enables broccoli head detection, pixel-level identification, determination of maturity categories, and precise localisation of suitable heads for harvesting, thus better aligning with practical harvesting scenarios. The maturity identification method is based on the DeepLabV3+ network model, which classifies pixel points into four categories: immature, semimature, mature, and hypermature. Furthermore, targeted enhancements to the network structure have been incorporated to accommodate the unique maturity characteristics of broccoli. MobileNetV2 contributes to the real-time detection of multiple broccoli heads within the view of camera. The Dense Atrous Spatial Pyramid Pooling (DASPP) module enhances the capability of recognising multiscale features of broccoli, and the Convolutional Block Attention Module (CBAM) further improves the integration of maturity information. The effectiveness of the targeted enhancements has been validated through ablation experiments. The semantic segmentation was successfully applied to broccoli maturity identification for the first time by incorporating a self-designed category determination module. The proposed algorithm achieves a mean intersection over union (mIoU) exceeding 57.9 %, the pixel accuracy (PA) reaching 98.56 %, and the mean category prediction accuracy (mCPA) of 70.93 %. These performance metrics outperform established algorithms such as BASNet, DeepLabV3+, and UNet. This advancement has resulted in an enhancement in the accuracy of maturity identification and a substantial reduction in computational expenses.