Comparing object recognition from binary and bipolar edge features

Abstract

Edges derived from abrupt luminance changes in images carry essential information for object recognition. Typical binary edge images (black edges on white background or white edges on black background) have been used to represent features (edges and cusps) in scenes. However, the polarity of cusps and edges may contain important depth information (depth from shading) which is lost in the binary edge representation. This depth information may be restored, to some degree, using bipolar edges. We compared recognition rates of 16 binary edge images, or bipolar features, by 26 subjects. Object recognition rates were higher with bipolar edges and the improvement was significant in scenes with complex backgrounds.

Figure 1

Image of ‘Watering pot’ (a) Color image of the watering pot in front of wire mesh and other clutter (b) Canny [7] binary edge result, 5 of 13 (40%) subjects identified the watering pot. (c) Bipolar edge [8] result. The bipolar edges and cusps act as a depth cue and may help segregate the object from the background clutter. Recognition rate was noticeably improved to 92.3% (12 of 13). Note that the images were scanned from prints resulting in artifactual outside edges interfering with the presentation.

Figure 2

Image of ‘Briefcase’ that was better recognized form binary edges than from bipolar edges (a) Original color image (b) Binary edge image calculated using the Canny detector. Average recognition rate was 53.9% (7 of 13 subjects). (c) Bipolar feature was detected by Peli’s method [8]. Due to the black color of outer briefcase, the depth from shading is inconsistently perceived in the bipolar feature. This might have reduced the recognition rate to 38.5% (5 of 13 subjects) although contrast polarity is consistent with the depth for numerous bright objects in the same scene (e.g. papers on the shelves).

Figure 3

Image of ‘Sprinkler’ that was much better recognized from bipolar images (a) Original color image (b) Binary edges image that no subject correctly recognized. (c) Bipolar edge image was recognized correctly in 53.9% (7 of 13 subjects) of presentations. The depth cues based on contrast polarity changes in the sprinkler might have helped to segregate the background and object.

Figure 4

Recognition rate for each object in all the test conditions. We divided Sanocki’s dataset into two groups based on their recognition results for counter balancing. As seen our subjects’ recognition rates were higher with the binary edge leaving little room for improvement with the bipolar processing (ceiling effect).