Improving the Structure of Deep Learning Algorithm in Image Processing Inspired by Representational Brain Dissimilarity Matrix

Deep learning algorithms achieves some results at human level or even better in pattern recognition problems. Meanwhile they apply a different mechanism other than human brain. This paper describes a human-inspired segmentation and interpolation algorithm, which applies the retinal layer in the proposed model after the input layer. Following this retina, this layer encrypts the input image and transmits the input image to the second space, which try to change deep network structure inspired of the brain's visual path. Network feedback, recognition rate, and network energy level or the comprehensiveness of the trained network examined in subsets of the Caltech data set. In similar examples, deep learning algorithms require more data to learn other than human. In the difference between deep learning and human, there is a difference in the representation of information. In deep learning, weights improve in a way that optimizes the result in a particular experiment, but in millions of years of human evolution, the human brain has evolved optimally and effectively representation. Another point of contention is the deepening of deep learning layers. The number of these layers has multiplied compared to the brain that lead to more complexity and energy expenditure. However, in the brain it can make a diagnosis with less energy. The maximum recognition rate of the proposed model is 93% and the base model is close to 91%. Also, the proposed model is thinner and the rate of fire of neurons in the initial layers is lower and has a high stability to changes in light intensity. The Dissimilarity of the model layers has been higher and it has been able to show a better response in the face of noise images and record less recognition loss.