A Salient Region Detector for structured images

Finding correspondences between two images of the same scene or object, taken from different viewpoints and in different conditions, is a challenging task. Furthermore, in the analysis of scientific imagery, it must be possible in terms of human perception to appreciate detected local features, thus making the task even more complex. A renowned generic feature detector, Maximally Stable Extremal Regions (MSER), performs very well on structured images, but has difficulties with blur, lighting and increased resolution. The detected regions do not always correspond to semantically meaningful image structures, and the large number of regions hampers scalability. This paper proposes a Data-driven Morphology Salient Regions (DMSR) detector which overcomes these limitations. We present a new binarization algorithm which uses a threshold derived from the data; the resulting binary image is analyzed for saliency using morphology. DMSR shows transformation invariance and comparable repeatability to MSER on several evaluation benchmarks while obtaining better invariance to lighting, blur and resolution. This is achieved via significantly fewer regions, leading to better scalability. Preliminary results on animal and plant images indicate that DMSR could be a suitable approach for wild-life biometric applications as the detected regions correspond well to the semantic image structures. We also introduce OxFrei - a dataset for transformation-independent detection evaluation.