Local-Feature Generic Object Recognition with Application to Insect-Species Identification

Abstract

Biological monitoring of stream water quality is a time-consuming and expensive activity for environmental protection and sustainable development. It requires intensive hand labor from experts to identify arthropod species known for their ability to reflect the status of their environment. The three generic recognition methods proposed in this dissertation are designed with the goal of developing image-based automated high-throughput insect-species identification with application to water quality assessment. These methods consist of a local-feature extraction and aggregation step to describe every image and a prediction step using this descriptor. The first method is based on the bag-of-features approach and uses unsupervised feature types and species-specific dictionaries to map sets of local features into histogram descriptors. These descriptors are concatenated and input to a logistic model tree (LMT) to discriminate among four stonefly species. The second method is a combination of efficient feature extraction and quantization using Haar-like features and random forests. The descriptors created by this process contain spatial information roughly correlated to specimen parts and are utilized by a spatial-pyramid kernel classifier for species identification. The final recognition method combines the classification scores of multiple types of local features while retaining their spatial information in a single spatial histogram of score accumulations. A stacked spatial support vector machine (SVM) is applied to these histograms in experiments with a set of 29 species matching those of a real biomonitoring task. The evaluation experiments performed demonstrate that automated insect identification is viable and can be both efficient and effective.