Object recognition in cluttered background is a challenging problem in computational modeling. When objects are present on natural backgrounds, the performance of object recognition models drop significantly. However, humans recognize objects accurately and swiftly despite this challenging condition. It seems that, our visual system achieves this ability based on lateral connections and feedback connections from higher areas.
One of the computational object recognition models that recently has achieved a remarkable performance in object recognition is convolutional neural network (CNN). It resembles feed-forward sweep of visual information processing. In this study, based on CNNs and inspired by biological evidence we proposed a recurrent object recognition model. The model simulates recurrent dynamics of visual object processing by implementing feedback and lateral connections. Evaluating the model to recognize objects on natural background, we showed that the proposed mechanisms significantly improves performance. In addition, visualizing the representations of layers indicatedthat deeper layers of the CNNs remove the background much better than the lower layers. According to the results, using both mechanisms -the feedback from higher layers and the interlayer surround suppression mechanisms- simultaneously in structure of CNN, the performance improvement was more than when either one was usedalone. This observation is in accordance withthe biological evidence from the human visual system.