COP Outcome development and evaluation

Initially, many options were considered, to improve spear size and weight estimation. One approach would be to actually weigh each spear after trimming to length, but existing, commercially available, weighing systems require the object being weighed to be stationary. This would require us to start and stop the line 12 times a second, which is physically impossible given the weight of the belt. The inertia forces required, even if available, would break the belt.

A parallel project into dynamic weighing of small bodies was done, giving promising results, but obviously requiring too long a time to complete. For these reasons, and because the existing machine had achieved fair results, we decided to stick with image processing as the base technology.

Firstly, it was established that the specific gravity of asparagus was very close to unity, and that density variation between varieties and from week-to-week during the harvesting season, was negligible at the level of control we were seeking. Once the density is regarded as fixed, the weight of a spear is directly related to its volume, so if the volume can be accurately estimated, an accurate estimate of weight follows. The reverse is also true.

The enormous increase in power and sophistication of computer systems over the last 10 years held out great promise of a big improvement in spear weight estimation and control of bundle weight.

The original designers faced many problems caused by the limitations of the hardware available to them at the time. For example, affordable video cameras had modest operating speeds and image resolution. The main problem, however, was limited computing speed.

When deciding on algorithms for weight estimation and spear allocation to bundles, their ingenuity was limited by the simple question “How much computation can we do in a twelfth of a second”?

This lead to them having to base their estimate of the weight of a spear on a single measurement, the diameter, measured two centimetres from the thick end. This is not as crude as it first appears, since all spears are eventually cut to the same length, but there is certainly scope for error. Firstly, the elliptical cross section of a typical spear is not constrained with the major axis in a known direction, so two identical spears could have significantly different measured diameters, and hence weight estimates. Secondly, spears differ in the amount of taper along the length, so two spears of identical real diameters could have different real weights, but would be assigned equal weight estimates.

COP Outcome development and evaluation

Experiments were done to establish a relationship between diameter and weight, and this relationship was used to convert measured diameter to estimated weight. Checks of actual versus estimated weight confirmed that the averages were equal, but there were quite wide errors for single spears. These were accepted as inevitable, but were a major contributor to the bundle weight problem.

The other main contributor was the algorithm which decided in which bundle to put the latest spear.

In general, at any one time there will be several part-complete bundles. The algorithm tried to complete a bundle, as its main objective. For example, if 4 bundles of “target – number = 5” were available, containing 1, 1, 2, 4 spears, the machine would try to put the latest spear in the bundle with 4. Only if this lead to a bundle weight of less than 100gram it would not do so, and put it in the one with 2 spears. It is clear that, since there is no check on the maximum weight of the bundle, overweight bundles will result. There are other, more subtle, knock-on effects of this algorithm which cause problems.

In the new system, the camera is able to capture an image of the whole spear as it passes a known point. On-board software downloads a digital version of this image to a computer which analyses the image. If the spear has any defects, such as white patches, extreme ovality, or is very bent, the computer allows the spear to run to the end of the belt and fall into a “reject” bin. If no defects are found, the computer calculates the projected area of the spear, which is converted by an experimentally determined equation to a volume, which leads via the known density to a weight estimate. These estimates have been compared to actual weight and found to agree much more closely. The standard deviation of the error is approximately a fifth of what it was previously. The conversion formula is updated at regular intervals; by ejecting a tiny fraction of spears off the line. They are automatically weighed and the information is fed back to the computer for corrective action.

The spear allocation algorithm was improved by adding the latest spear to the bundle where it would make the average weight of spears in the bundle most closely match the nominal weight. For example, if the 100 gram bundle has a target number of 5 spears per bundle, the average weight of a spear must be 20 grams. If a bundle currently holding 3 spears weighs 62 grams, a spear of 18 grams, when added, will result in 4 spears weighing 80 grams, a perfect average of 20 grams each. This method checks for both under and over weight bundles, and thus achieves bundles much closer to the target weight, and with even less chance of being under weight.

None of these improvements could have been implemented at the required speed of 12 spears/second, without the benefit of the latest technology:

• High speed video cameras.
• Sophisticated image processing programs.
• Rapid computation of weight estimates.
• Rapid allocation of spear to bundle via sophisticated algorithms.
• Reliable electro-pneumatic control equipment for spear and
  bundle handling.