Detecting fruits in natural scenes by using spatial-frequency based texture analysis and multiview geometry

This paper describes a computer vision based model for object detection that can serve as a preliminary step in fruit prognosis, which involves the estimation of the number, diameter and yield of apple fruits. In order to overcome the recognition unreliability in uncontrolled environments caused by uneven illumination conditions, partly occluded surfaces, and similar background features, we rely on a combination of the object's colour, texture and 3D shape properties. In our research, we apply colour segmentation to multiple scene snapshots to separate potential regions from the background and verify them first with texture analysis and second by reconstructing them to 3D space. By analysing all three distinct features (colour, texture and 3D shape) of possible areas, we can safely conclude if they represent fruits we are looking for. Once we detect and verify all areas representing fruits, we can measure their size and model estimated fruit yield.

Research highlights▶ The paper describes computer vision based model for object detection. ▶ We examine three distinctive features: colours, texture and 3D shapes. ▶ For texture analysis Wigner–Ville distribution and support vector machines are applied. ▶ Projective reconstruction is used to reconstruct 3D shapes from 2D views.