Development of an advanced synthetic data modelling engine capable of automatically producing high volumes of variable complex crop scene datasets at >10% of real-world costs within days rather than months/years to train agricultural robots

农业机器人需要经过有效训练的人工智能系统来有效地执行功能。农业领域是最难训练人工智能系统来解释农业场景的领域之一，这是由于多层次的复杂性：* 植物/杂草物种：** 巨大的物种差异和多个物种在早期生长阶段都很难区分 * 遮挡：** 在复杂的作物场景中，许多作物和杂草植物以复杂的方式重叠 * 物理变化：** 害虫/疾病、叶子/作物畸形和土壤变化的影响 * 不同的呈现：** 相机角度、场景照明和背景会产生变化和模糊效果 * 注释：** 人类几乎不可能在像素级上精确地对图像进行大量注释英国农民面临的最困难和经济上最具破坏性的问题之一是黑种草，它威胁到小麦作物的生存能力。对于农业机器人来说，在早期阶段很难检测/区分黑种草，这是有效消除处理所需的，因为这需要大量的、变化的数据集，可能需要数年时间才能获得。在这个项目中，我们将开发一个先进的合成图像建模引擎，能够以低于现实世界成本的10%自动生成大量可变复杂作物场景数据集。这些可以用来在几天内有效地训练AI解决方案，而不是几个月/几年。这将使农业机器人能够稳健地检测黑种草。黑种草是第一个用例，该项目的进一步发展将使其他物种能够在广泛的条件下进行分类。

Agricultural robots require effectively trained AI systems to carry out functions effectively. The agricultural sector is one of the most difficult in which to train AI systems to interpret agricultural scenes due to multiple layers of complexity:* **Plant/weed species:** huge species variances and multiple species both difficult to distinguish at early growth stages* **Occlusion:** In complex crop scenes many crop and weed plants overlap in a complex manner* **Physical changes:** Effects of pests/diseases, leaf/crop deformities and soil changes* **Different presentations:** Camera angles, scene lighting and backgrounds create variabilities and translucency effects* **Annotation:** Annotation of images at pixel level is almost impossible for humans to do accurately and at volumeOne of the most difficult and economically damaging problems for UK farmers is blackgrass, which threatens the viability of wheat crops. Blackgrass is difficult for an agricultural robot to detect/distinguish at the early stage which is required for effective elimination treatments as this requires a significant, varied, dataset which could take years to obtain. Such a robust AI solution does not exist today.During this project, we will develop an advanced synthetic image modelling engine capable of automatically producing high volumes of variable complex crop scene datasets at <10% of real-world costs. These can be used to effectively train AI solutions within days rather than months/years. This will enable agricultural robots to robustly detect blackgrass.Blackgrass is the first use case and further development within the project will enable other species to be classified in a wide range of conditions.