Object detection based on cortex hierarchical activation in border sensitive mechanism and classification-GIou joint representation

By imitating the brain neurons for object perception, the deep networks enable a comprehensive feature characterization in the task of object detection. Considering such a perceptual ability is usually bounded in a box area for feature extraction, the balance of dimension reduction and feature information retaining has been taken into account in more recent studies, especially for the information preservation in the border areas. Motivated by the mechanism of neuron cortex activation, in this work, a novel function based on cortex hierarchical activation is proposed to achieve more effective border sensitive mechanism by joint pooling in backbone networks. In order to avoid the parameter solidification, this strategy is also capable to benefit the feature extraction on the border without unnecessary model re-training. Furthermore, by replacing the square kernel with a designed band shape kernel, more adequate feature description can be obtained on the border via the combination of the strip hierarchical pooling and strip max pooling. With an extension of the proposed activation function on classification-GIoU joint representation, the overall detection accuracy has been further improved. Experimental evaluations on the COCO benchmark datasets have shown that the proposed work has a superior performance in comparison to other state-of-the-art detection approaches.