In a first step, an assessment protocol of classification systems was developed and simulated for our particular problem. The evaluation measures the different sorts of errors (incorrect reject, incorrect detections) and their statistical validity. Investigating the recognition unit consisted in developing and choosing relevant image descriptors for species based discrimination, and using several classifiers designed with neural networks. A morphologic analysis of fishes made it possible to free the process from knowing their position in front of camera. Simulations help to assess different classification strategies from images acquired by the machine without the cost and the limits of a full-scale test (installation of equipment at Lorient auction market, reception of fresh fishes when boats unload their catch).
Acsystème designed a neural network learning software to set up strategies according to an image database of which the purpose is to compare strategies to each other on the basis of costs related to a compromise between conceivable errors (incorrect reject against incorrect detection).
Finally, Acsystème implemented the simulation code into the machine’s software environment in order to dynamically load learned neural networks for new species recognition.
The neural network represents the computer programs’ artificial intelligence but numerous elements limit their capacities:
- The more complex the problem is, the more important the size of the database is, increasing the cost in terms of time and money.
- Descriptor definition needs an expert’s analysis. With the same database, different definitions of descriptors may generate neural networks with very different performance levels.