Object Detection Model Based on Dual Visual Pathway

Object detection has made great progress with the development of deep learning. However, existing object detection algorithms still have limitations compared with human brain. In addition, the interpretability of the model becomes more and more inefficient with the increasing accuracy of the model. Most of the object detection models are considered as "black box" that cannot be explained. Considering that the visual information processing system is composed of two pathways spatially distributed in the ventral and dorsal sides of the cerebral cortex when the brain detects objects, this paper proposes object detection model based on dual visual pathway by simulating the process of detecting objects in the brain. Intra layer recursion and remote connections of different modules are added to the feedforward network to achieve feature extraction for image, since the ventral flow pathway is a recurrent network composed of loops and feedback. Experimental results carried out on PASCAL VOC show that the performance of object detection has been further improved, and the interpretability of the model has been enhanced.