Research On The Key Techniques Of Image-based Insects Recognition

Insects recognition is the basis of crop pest and disease control in agriculture and forestry. Traditionally experts observe the external features of insects and then compare these features with specimens while identifying insects, which is time-consuming and labor exhaustive. Existing methods for insects automatic identification is simple functionally and their performance needs to be improved as well. Under the support of National Nature Science Foundation Project - Research on Data Mining Based on Medical Image and the support of Shannxi's Province Nature Science Foundation Project - Research on Semantic Annotation and Ontology Retrieval of Relics, geometric model-based image segmentation, content-based image retrieval, dimension reduction-based recognition, ontology building and ontology-based recognition as well as several other key techniques are studied according to different stages of identification i.e. pre-processing, extraction, classification in image-based insects recognition in the thesis. This research is of theoretical and application value. Main contribution of this thesis is as follows:(1) New hierarchy mechanism for insects recognition is constructed. Insects recognition is divided into two levels that are low-level visual features based recognition and high-level semantic-based recognition. Through the mapping establishment between low-level visual features of insects and the high-level semantic, semantic gap between these two levels are narrowed so that ontology-based insects recognition is implemented.(2)Visual features-based insects recognition. The features of insects images are extracted according to texture features, local features, reduced-dimension technology from three different angles and then these features are classified by corresponding classification algorithms, so that better performances are gained by these methods.a)Image segmentation preprocessing methods for color image of insects are presented. Because the existing geometric model can only be used for deformation gray image segmentation, rapid geometric model of deformable segmentation is presented and implemented so that color images can be well dealt with.b)Angle invariant Gabor based SVM for insects recognition. The texture features are extracted by Angle invariant Gabor and represented as texture features matrix, and then SVM is applied on the matrix to recognize insects.c)SURF based Multi-resolution histograms for insects recognition. According to different light condition and scale, local features techniques are introduced to insects recognition. Based on the invariant of local features to light condition and scale, SURF descriptor is used to extract local features of insects images. Since the number of local features are unfixed, multi-resolution histogram is applied to local features matching.d)Spectral regression's LDA-based KNN insects recognition. Since insects recognition in high-dimensional space usually leads to dimension disaster, spectral regression based LDA dimension reduction is adopted so that insects are projected into certain features subspace, and then KNN classification is implemented to insects recognition.(3)Ontology-based insects recognition. By taking advantage of the mapping between visual features and insects type, ontology engineering approach is adopted to implement insects recognition.a)Establishment of ontology of insects morphological taxonomy. Since the knowledge hierarchy of insects taxonomy is complicated, insects classification metadata knowledge units are extracted and mature ontology establishment methods are adopted to achieve formal and sharing knowledge of insects.b)New mechanisms of ontology-based insects recognition is presented. Insects visual features ontology, media feature ontology are established. By combining these ontologies, and taking advantage of the mapping between insects' visual features and high level semantic, ontology-based insects recognition is implemented.(4)Insects Recognition system demo is developed. Using protege to establish ontology, in combination with Visual C++ 6.0 and Matlab, the algorithms concerned in the thesis are implemented and integrated. The correctness and precision of algorithm are tested and the effects are good.