With the growing capabilities and the spreading of computer vision tools into various industries it is inevitable that we will see more and more of these tools being used in the fields of agriculture.
One such field is intra-row cultivation, also known as within-row or in-row cultivation is the cultivation of the crop plants within the field rows, and until the advent of computer-controlled and vision guided equipment was a costly and tiresome job that needed to be done by hand.
The basic premise when approaching this problem from an image processing point is that we know to recognize the rows themselves (as was discussed in our article about inter-row cultivation) and want to either recognize the plant itself or conversely recognize the weeds within the row.
The classic approach would have been to model the plant in question (or, more rarely, the weed in question) and attempt to recognize it in the images. This approach suffers of course from the dual curses of being too restrictive (as it is an extremely difficult to devise a model that will effectively screen out weeds while being inclusive enough to allow a wide enough variation in the crop plant to be effective) and unscalable (as we will have to devise such a model for each and every crop plant we wish to work with).
A solution to this problem could come from looking at the weeds as disturbances in the order imposed by the repetitiveness of the plants in the row while ignoring the small local plant variations. While sounding impressive this can be as simple as learning the intensity frequency in every row and using the assumption that while crop plant appear in the row in a consistent frequency – the weed distribution is random. This solution can be enhanced by capturing as many rows as possible with every frame thus allowing us to validate our results between the rows.