Research On Deep Networks And Image Detection And Recognition Inspired By Biological Vision

With the emergence of big data such as images and videos and the improvement of computing capabilities of GPU components,the development of deep learning technology and theory has been promoted.Computer vision has also obtained a new development opportunity.The applications of image classification and detection based on artificial intelligence are also becoming more widespread.Study on image classification and detection has become a research hotspot in computer vision.The existing methods have demonstrated good performance for image classification and detection.There is always a delicate relationship between computer vision and biological vision.Convolutional neural network(CNN)is able to learn features and memorize the training data,which can be considered as a simulation of the information processing mechanism of human vision.However,it is a challenge to reduce computational complexity and the number of model parameters.In the thesis,we apply the correlated theories and mechanisms of biological visual system,and bring forward some novel bio-network computing model and some intelligent algorithms.The main work and results are as follows:(1)We propose a biologically inspired visual integrated model for image classification called VMVI-CNN.Motivated in part by recent neuro-science progress in revealing integrated functions of human visual system,two models of bio-inspired visual mechanisms(visual memory decay mechanism and visual interaction mechanism)are proposed and built within the VMVI-CNN to(1)control the feature information passing through,and(2)increase the rich-ness of feature information.The proposed method is tested on three benchmark datasets and a real-world industrial dataset.The results demonstrate that the new model can extract distinctive features and exhibit a better recognition performance than the current state-of-the-art approaches(2)As the critical machine learning method,CNN was proposed according to the hierarchical structure of biological vision,which could achieve superior performance in large-scale image classification.In this paper,inspired by biological mechanisms and structures in neuroscience,we propose a new framework called visual interaction networks(VIN-Net),which is inspired by visual interaction mechanisms.More specific,self-interaction,mutual-interaction,multi-interaction,and adaptive interaction are proposed in VIN-Net,forming the first interactive completeness of the visual interaction model.To further enhance the representation ability of visual features,the adaptive adjustment mechanism is integrated into the VIN-Net model.Finally,our model is evaluated on three benchmark datasets and two self-built textile defect datasets.The experimental results demonstrate that the proposed model exhibits its efficiency on visual classification tasks.(3)As the critical machine learning method,CNN has shown superior performance in image classification and other computer vision tasks.However,it is a challenge to reduce computational complexity and the number of model parameters.In this paper,we propose a new model called partitioning CNN,which is inspired by visual local learning mechanism.Through the use of biological inspired partitioning CNN,the new model can reduce the computation cost,extract distinctive features,and have a good recognition performance.The proposed model is tested on two benchmarks and a realworld industrial dataset.The results demonstrate that the model has better performance and lower GPU consumption than the current state-of-the-art approaches(4)Finally,based on the related biological mechanism and biologically inspired deep networks,we proposed a practical application scheme by biologically inspired deep network models.In particular,in the context of textile industry,we propose a variety of algorithms for textile defect classification and detection.Furthermore,we conduct extensive experimental validations for various design choices.The experimental results show that the proposed methods outperform the existing classification and detection methods.