Filtering by
- All Subjects: Machine Learning
- Creators: Bansal, Ajay
- Status: Published
In this thesis, several different methods for detecting and removing satellite streaks from astronomic images were evaluated and compared with a new machine learning based approach. Simulated data was generated with a variety of conditions, and the performance of each method was evaluated both quantitatively, using Mean Absolute Error (MAE) against a ground truth detection mask and processing throughput of the method, as well as qualitatively, examining the situations in which each model performs well and poorly. Detection methods from existing systems Pyradon and ASTRiDE were implemented and tested. A machine learning (ML) image segmentation model was trained on simulated data and used to detect streaks in test data. The ML model performed favorably relative to the traditional methods tested, and demonstrated superior robustness in general. However, the model also exhibited some unpredictable behavior in certain scenarios which should be considered. This demonstrated that machine learning is a viable tool for the detection of satellite streaks in astronomic images, however special care must be taken to prevent and to minimize the effects of unpredictable behavior in such models.
Kitsune attempts to remedy these issues by tying itself to Antlr, a pre-existing language recognition tool with over 200 currently supported languages. In addition, it provides an interface through which generic manipulations can be applied to the parse tree generated by Antlr. As Kitsune relies on language-agnostic structure modifications, it can be adapted with minimal effort to provide plagiarism detection for new languages. Kitsune has been evaluated for 10 of the languages in the Antlr grammar repository with success and could easily be extended to support all of the grammars currently developed by Antlr or future grammars which are developed as new languages are written.
The main focus of this thesis is to use visual description of a landmark by choosing the most diverse pictures that best describe all the details of the queried location from community-contributed datasets. For this, an end-to-end framework has been built, to retrieve relevant results that are also diverse. Different retrieval re-ranking and diversification strategies are evaluated to find a balance between relevance and diversification. Clustering techniques are employed to improve divergence. A unique fusion approach has been adopted to overcome the dilemma of selecting an appropriate clustering technique and the corresponding parameters, given a set of data to be investigated. Extensive experiments have been conducted on the Flickr Div150Cred dataset that has 30 different landmark locations. The results obtained are promising when evaluated on metrics for relevance and diversification.