For a customer in the agricultural sector, a computer vision system will be developed which observes an organic product on a conveyor belt and produces a qualitative description of the product.

• Preliminary study of the effect of different types of (visible, IR, UV) lighting which helps the detection of small blemishes/deceases.
• 3D object modeling to produce an accurate size/weight estimation
• Texture and color analysis
• Hardware integration, allowing a robotic arm to grab the product from a fast-moving conveyor belt.
• Feedback from a weight sensor in combination with a self-learning algorithm to produce accurate weight estimates.

A recent project has demonstrated the effectiveness of AI driven computer vision for agricultural applications.

• Crops (or any kind of organic substance) can be reliably detected, measured and classified/categorized. AI algorithms are surprisingly good at handling problem spaces with fuzzy or ill-defined boundaries. For example, no two potatoes are exactly alike and the quality of a potato is determined by a number of subjective criteria. Nevertheless, smart algorithms could reliably detect, model the 3D shape and classify the quality.
• Pre-harvest analysis is possible for several types of crops. This may allow you to better plan/schedule the harvest and increase your revenue.
• During the entire growth process a large number of variables will determine the final harvest results, many of which can be controlled (humidity, temperature, air flow, fertilizer). Determining the best settings for all these parameters at any point in time is something that’s often done bases solely on experience and “feeling” with the process. AI algorithms can capture this experience and then improve on it by self-learning from results, which in the end inevitably leads to a more optimal system.
• Post-harvest analysis using smart algorithms can improve the sorting of your product; shape, color, texture, bruises, deceases can all be detected as well as a trained worker could do it, in many cases an automatic sorting could be better due to the additional sensor possibilities that are available to a computer (IR, UV, textured light, etc).

We nowadays make extensive use of automatic object classification based on a very limited amount of information. Sorting food items by size, weight or even color is not all that hard to automate. However, removing the items that are bruised or cracked and selecting premium food items based on surface patterns or shape is a task that is barely done in another way than manually although it is made possible with certain vision techniques.

Manual inspection for quality control is a tedious and labour intensive process. Contrary to common belief, automatic visual inspection can not only reduce labor costs, but can also increase the inspection quality.

Examples of object inspection in an industrial or agricultural setting:

• A self learning system can be trained to inspect the looks of strawberries, analyzing the color, shape, checking for irregularities and finally giving an accurate quality rating. Automatic inspection allows you to make thousands of inspections per minute, easily outperforming a human observer. (more info)
• Inspection of buildings and infrastructures for signs of cracks or wear. Combine with a robotic solution to inspect places that are dangerous or hard to reach for a human observer, like air vents or sewer systems.

Train a system yourself to new situations or products simply by presenting positive and negative examples. A self learning system can continue to learn and adept itself, whereas traditional systems would constantly require expert consultancy.