Intelligent Computing Model Inspired By Biological Vision And Its Application

The human biological vision system is considered to be one of the most advanced biological intelligence systems,and it plays a vital role in the process of human interaction with the environment.In recent years,more scholars have begun to pay attention to the internal computer theory of the efficiency,plasticity and mobility of the biological vision system.At the same time,the derivative theories inspired by biological vision such as deep learning have become more and more hot research issues in the field of image processing.one.Inspired by biological vision,this topic comprehensively considers traditional biological vision mechanisms and cutting-edge deep learning theories,designs visual intelligent computing models,and uses traditional edge detection,current popular small sample learning,and image defogging as examples to carry out research.The main research contents of this topic are as follows:(1)Inspired by the complementary characteristics of long and short-term synapses in the work of biological visual perception systems,an edge detection method based on long and short-term synaptic complementary networks is proposed.Construct a neuron network with complementary characteristics of long and short duration synapses.First,introduce the dominant color antagonistic characteristics of the cone cell population,and perform weighted coding on the color antagonistic channel of the image to be tested to obtain the primary edge perception of the image to be tested;then simulate;Synchronous discharge characteristics of neuron group,define the neuron action window of synaptic dynamic connection,realize the group discharge time coding for primary edge perception;then construct the long-and short-term synaptic complementary module,based on the short-term synchronization discharge characteristic of neuron group And long-term neuronal firing activity time sequence and space dependence,to achieve long-term and short-term synaptic plasticity coding and complementary fusion,and finally obtain the edge response by encoding the time information flow.The 20 colony images collected by this laboratory according to the requirements of routine microbial experiments are used as experimental materials,and the reconstruction similarity MSSIM,edge confidence BIdx and comprehensive index Eidx are used as evaluation indicators.The results show that,compared with the three mainstream methods of VSC,NIS and MSP,the detection results of this research algorithm are accurate,and the missed detection rate is low,which is more consistent with the manual subjective observation results;meanwhile,the mean and standard deviation of the EIdx index are 0.804 8 ±0.052 1.The overall performance is better than the above three mainstream methods.(2)Inspired by the small sample dependence of the human biological visual system,a small sample learning method based on visual hierarchical feature coding is proposed.Considering that similar small sample images have commonalities in low-level features such as direction and color,we design low-level feature coding areas based on the orientation perception characteristics of the primary visual cortex and the color perception characteristics of cones,and extract low-level features such as the direction and color of the image to be recognized;Then,the deep residual module is designed by using the hole convolution feature enhanced by redundancy to analyze the high-level semantic features of the image;finally,the cosine similarity is used to measure the similarity between the object to be recognized and its category by means of metric learning.The Omniglot data set and mini Imagenet data set are used as experimental materials,and the model training is carried out with the experimental methods of "5-way and 1-shot,5-shot,and 10-shot" respectively.The experimental results show that the recognition accuracy of the three combinations in the Omniglot dataset and mini Imagenet dataset reaches 98.8%,99.2%,99.4% and 55.39%,77.24%,78.31%,and the loss convergence speed is extremely fast,saving time And resource costs.(3)In order to verify the generalization ability and algorithm effectiveness of the intelligent computing model designed in this paper,the proposed model is applied to image defogging.This paper first considers that different levels of image features play different roles in the semantic expression of the image,and design a hierarchical encoder to extract the low-level features and high-level semantic features of the image respectively;then,use separable shared convolution to improve the traditional hole The risk of local spatial information inconsistency in convolution;then,through the active fusion of low-level features in the feature processing link,the image features of different levels are aggregated to achieve image enhancement;finally the enhanced image features are decoded and combined with the foggy image Fusion,get the final image dehazing result.The OTS data set is used as the experimental material,and the structural similarity SSIM and the peak signal-to-noise ratio PSNR are used as quantitative evaluation indicators to evaluate the effectiveness of the algorithm from both qualitative and quantitative perspectives.The results show that compared with the five methods of Meng,DCP,NLD,DCPDN and DHN,the algorithm in this paper effectively removes the fuzzy features of the foggy image,and clearly restores the color and true shape of the image.At the same time,the average value of SSIM index is 21.0113,and the average value of PSNR index is 0.8175,which is better than the above five mainstream methods.