Using sparse coding as a preprocessing technique for insect detection in pulsed LIDAR data
Zsidisin, Connor Reece
MetadataShow full item record
This research proposes using sparse coding as a preprocessing technique on insect lidar based data. This preprocessing technique will be used in conjunction with the Adaptive Boosting (AdaBoost), Random UnderSampling Boosting (RUSBoost), and neural network algorithms to automatically detect insects. The project aims to increase the effectiveness of these algorithms by using new images created by sparse coding. The K-Singular Value Decomposition (KSVD) algorithm will be used to train a dictionary on images that contain the majority class (non-insects). This trained dictionary will be used along with Orthogonal Matching Pursuit (OMP) to reconstruct all lidar images. The difference between the original image and the reconstructed image will be taken and processed by the feature extraction function and then used to train and test the models. Using a complete and an overcomplete dictionary our results show that the algorithms are able to detect insects at a higher rate. Using an overcomplete dictionary we are able to classify 93.18% of insect containing images in the testing dataset. Using the complete dictionary we were able to maintain 99.70% of non-insect images while increasing the percentage of insects classified to 84.09%.