Speaker
Description
Reliable end to end traceability of individual sawlogs from the forest to the sawmill is a key enabler for digitalisation, process optimisation, and value recovery in the forest value chain. Code based log identification using computer vision offers a scalable alternative to physical tagging, but its practical deployment is challenged by variable visual conditions, partial occlusions, and classification uncertainty typical of forest and industrial environments. This study investigates the robustness of Reed–Solomon–based symbol encoding for code based sawlog identification under realistic detection and classification noise.
We present a simulation based evaluation of a computer-vision classifier system using an Reed-Solomon RS(8,4) symbol code to encode log identifiers. The code consists of four data symbols and four parity symbols, providing a minimum Hamming distance of five and enabling correction of multiple symbol errors or erasures. To emulate realistic classifier behaviour, synthetic confidence matrices are generated by jointly modelling per symbol detection probability and inter class discriminative power. This framework enables controlled analysis of identification performance across a wide range of noise levels, symbol loss rates, and dictionary sizes, up to the full RS code space of 23⁴ possible identifiers.
Code recovery performance is evaluated using both hard decision decoding, which selects the most likely symbol at each position, and soft decision decoding, which exploits the full confidence distribution across symbol candidates. Results show that soft decision decoding consistently and substantially outperforms hard decoding across all tested scenarios. In low noise conditions, correct identification rates exceeding 99% are achieved even for per symbol detection probabilities as low as 0.8, without applying any dictionary constraints.
Under moderate noise, identification rates above 98% are maintained for dictionary sizes of up to 20 000 identifiers, demonstrating strong scalability for industrial scale deployments. At high classification noise levels, and symbol detection probabilities as low as 50%, performance becomes increasingly dependent on both code design and dictionary filtering. Here, 74% of identifiers are correctly recovered using dictionary-constrained decoding, primarily due to the strong minimum distance properties of the RS(8,4) code. This performance is primarily attributable to the code’s minimum Hamming distance of five, which guarantees unambiguous recovery in the presence of up to four missing or erroneous symbols. Notably, even when more than four symbols are missing or misclassified, the dictionary-constrained decoder frequently identifies the correct code candidate, highlighting the benefit of combining error correcting code structure with probabilistic decoding.
Overall, the results demonstrate that Reed–Solomon–based symbol encoding, combined with soft decision decoding and optional dictionary constraints, enables robust and scalable code based sawlog traceability under challenging visual conditions. The proposed approach supports reliable digital identification of individual logs in forest and sawmill environments without reliance on physical tags or manual intervention.
| Keywords | forest digitalization; computer-vision; deep-learning |
|---|
