Research On Small-Scale Image Recognition Based On Convolutional Neural Network

The advent of the era of big data and the rapid improvement of computer computing capabilities has promoted the development of image recognition technology in a more advanced direction.Image recognition technology based on deep learning has become a research hotspot in the field of artificial intelligence.Convolutional neural networks,as one of the deep learning algorithms,have been widely used in the field of image recognition due to their superior performance.Compared with the traditional image recognition method,the convolutional neural network no longer needs to manually select features.It can self-learn and characterize the image information,and the obtained image recognition rate is even flat or even beyond the human eye recognition level.In the context of today's big data era,convolutional neural networks are clearly a good choice for the identification and classification of massive images.Convolutional neural networks can have hundreds of millions of parameters,such as the convolutional neural networks AlexNet,VGGNet,Google Net,etc.,which perform well in large-scale visual recognition challenges.These networks need to train a large number of parameters,not only requiring high computing platforms,but also requires a lot of image data.It is unrealistic to train these complex networks directly on small-scale image datasets such as Caltech-256 and Food-101,which do not provide enough parametric learning space for these large-scale networks.Although these networks can be applied to the identification of small-scale images by means of migration learning,these network layers are often deep,the network is complex,and the number of parameters is relatively large,which cannot be deployed in some environments with lower hardware platforms.Aiming at this problem,this paper designs convolutional neural networks that can balance the size and accuracy of the model.It is suitable for the identification of small-scale images and is convenient for deployment in some low-level environments.The main contents and innovations of this paper are as follows:(1)Summarize the optimization methods of convolution neural network and prevent over-fitting technology,analyze the design methods of classical convolution network,and expect to use simpler network parameters and model size to achieve the accuracy level of similar methods in small-scale image recognition,even surpass the same methods.(2)Design a lightweight and ultra-lightweight convolutional neural network to evaluate the performance of the designed network on selected small-scale image data sets and compare it with the existing methods to illustrate the advantages of the designed network in small-scale image recognition.(3)Based on the classical AlexNet,a parallel concatenated convolution network based on cross-layer connection is designed for small-scale image recognition.Parallel convolution is to extract features from convolution kernel of different scales in parallel,and to fuse features extracted from different scales in parallel.In order to further improve the accuracy,the network performance was evaluated on the Caltech-256 and Food-101 datasets by adding a cross-layer connection to optimize the network.The results show that compared with the classical AlexNet,RPCNet based on cross-layer connection improves the accuracy by 6.12% and 12.28% respectively,and the network size is only 1/15 of ALexNet.(4)A lightweight convolutional neural network architecture GResNet based on grouping residual structure is designed for large-resolution small-scale image data sets.The core of this method is to divide the upper output feature map into four groups with equal number by using the bottleneck structure of three convolution layers.Residual mapping is added to the bottleneck module in the group and the adjacent module outside the group respectively,so a lightweight convolution neural network is designed.In the experimental stage,the performance of the network is evaluated on the small data sets Caltech-256,Food-101 and GTSRB.The experimental results show that GResNet has the same or even better recognition performance with fewer network parameters than traditional convolutional neural networks,and is suitable for deployment in low hardware platforms.(5)Ultra-lightweight convolution neural network architecture is designed for small-scale image recognition with small resolution.The method is to optimize the baseline network BasictNet by using the designed two-dimensional decomposition convolution and deep separable convolution structure.In the experimental phase,network performance was evaluated on Cifar10 and Fashion-mnist datasets.The results show that the accuracy of convolution network TwoNet based on two-dimensional decomposition is 1.08% and 0.32% higher than that of baseline network BasictNet,and that of deep separable convolution network DwNet is 1/3 of that of BasictNet,but it only loses 1.43% and 0.32% accuracy on two datasets,and the network model size is less than 2M.Relatively,removing the deep separable operation of multi-scale network DwNet+Ms can improve the accuracy by 1.35% and 0.56%.Compared with some existing network models,these networks can achieve the recognition accuracy of some advanced networks with less than 1M parameters,and reduce the scale of the models by more than 100 times,which has certain application value.