An Adaptive Multi-Level Framework for Forest Species Recognition

In this work we propose an adaptive multi-level approach for combining multiple classifications, applied to forest species recognition. This approach is based on the idea that the recognition of an input image can be improved by extracting from it multiple feature vectors, which can be by means of multiple segments and/or multiple feature vectors, and by combining their classification results. Generally, the higher the number of feature vectors, the higher the accuracy. On the other hand, by handling a smaller number of vectors, less classifications are performed and the recognition task can be done faster. In order to avoid the burden of evaluating a high number of feature vectors for any image, the proposed adaptive multi-level framework increases the number of feature vectors sequentially depending on the test sample. In this case, the more complex but more accurate approach is used only for images which the classification scheme does not yield enough confidence in previous levels. In the experiments we demonstrated that the proposed framework is able to attain similar performance of a static single-layer system but with considerable reduction in complexity (up to 4/5).