Ports haul in clever neural nets

Getting containers out of superports must be achieved as efficiently as possible. A neural net system can help out considerably

In the giant superports of today, where business is won or lost on a few pence per container, it is imperative to get containers in and out of the port as quickly as possible.

A highly efficient tracking system is obviously needed. Yet up until now, the observation part of the system has best been handled by people. Although a lot of time is spent tracing human errors, or looking for containers that have been mislaid, this is still more effective than using modern optical character recognition systems (OCRs).

That is because of the difficulty of recognising data on the containers. Containers may have arbitrary colours, surfaces may be corrugated, dented or rusting, lorries may be moving, or there may be little light. Compounding the issue, there are no standard formats or fonts for the characters on the containers, and even the most sophisticated OCRs on the market cannot manage more than about a 60% correct read rate.

Cambridge Neurodynamics, now has produced a neural network container tracking system which has a claimed 90% read rate in hostile port environments. Running on standard PC equipment, the system is not programmed, but trained on a large number of sample characters, learning to cope with poorly defined characters in the video image.

It consists of three components: cameras, an image capture board (which takes images from a standard video input), and a parallel processing board which contains a neural network recognition engine.

The networks were trained using a live video of passing containers at a working port, allowing the system to learn to cope with the variety and damage encountered in practice.

As containers pass multiple cameras, the unit captures multiple images. Letter finding is performed by advanced self learning methods, which localise areas of image containing letters. Containers often have non-relevant lettering and series of areas rich in letterlike artifacts.

The Cambridge system isolates possible areas of interest. Candidate areas are passed to recognisers, one for each symbol. Areas with letterlike characteristics but no letters are quickly rejected. The position of the recognised characters and their values are input to a format detector which is capable of rejecting non-relevant writing on the container.

Figure 1: Example of a container tracking system configuration to track container, driver, damage and vehicle number plate

{{Cambridge NeurodynamicsTel: Cambridge (01223) 421107Enter 440}}