Description

ABSTRACT:

•      Our Proposed method  allows for the manipulation of two crucial variables for image classification: discriminability and invariance. Our filters are modeled and trained to optimally fit local image structures and as a results, generate a good balance between discriminative representations and invariance. In particular, our results on three natural image sets and one synthetic image set highlight the increase in discriminative power of our network as well as its robustness to local geometrical transformations. Moreover, spatial organization of local features into a global representation is a key aspect to image recognition. In this regard, the multi-resolution pooling introduced in this paper provides rich spatial information, resulting in improved classification scores.