Small Sample Image Augmentation Method Based On Particle Swarm Optimization Algorithm

Image classification is a research direction in computer vision today.As a main research method,deep learning has a strong function fitting ability due to the deep complexity of the network architecture,but for small sample image data sets,there is still an overfitting problem.To s olve the over-fitting problem,you can optimize the network structure and i ncrease the number of t raining i mages.N owadays,a l ot of research is de voted t o t he design and optimization of the network structure,and less attention is paid to finding the method of image augmentation.F or e xample,the ba sic image a ugmentation method us ed on t he I mage Net dataset proposed in 2012 is still used as the standard for data preprocessing.These basic image processing methods are only effective for specific data sets,and are less robust when applied to other data sets.Compared with collecting more actual image data,image augmentation does not require a lot of manpower and material resources,so it is widely used in the task of object classification a nd r ecognition.I n t he pa st,the implementation of i mage a ugmentation w as based on the basic processing of the image.These augmentation methods only assumed that the image had basic characteristics such as rotation,translation,and scale scaling invariance,and had nothing to do with the data content in the specific image.Nowadays,when the model is complicated and the computer computing performance is bottlenecked,image augmentation,as a means to break through the classification performance,has its research and development necessity.Aiming a t the c haracteristics of s mall s ample i mages w ith s mall a mount o f d ata a nd imbalance,th is p aper models the i mage a ugmentation p roblem f rom the perspective of optimization,and models it as an optimization problem with the optimal distribution in the image f eature s pace.By an alyzing t he o ptimization g oals an d co nstraints,an i mage augmentation method based on the iterative idea of particle swarm optimization algorithm as the pr ototype i s pr oposed,a nd t he p ixel s pace o ptimization i s pe rformed on t he a lgorithm prototype according to certain rules and constraints.The demand hypothesis of the research is that the main characteristics of the image category to be augmented generally obey the normal distribution.The optimization goal is that the normal distribution of the image feature space satisfies t he i nvariance o f the m ean and the i ncrease of the s tandard deviation.The s earch strategy is a co mbination of two meta-strategies.The decision space of the algorithm is the feature space,and a set of solutions in this space and its corresponding images are optimized.The algorithm designs an augmentation strategy according to the characteristics of the image type,and uses the result of intermediate iterations as a n ewly generated image to construct a more ba lanced data distribution a nd a lleviate t he ove rfitting pr oblem o f t he deep ne ural network.In the experimental part,image augmentation is carried out on s ome representative categories in the omniglot d ata s et and min i-imagenet d ata s et.C ompared w ith t he state-of-the-art image augmentation and enhancement methods by controlling variables,t he algorithm in this paper can obtain a wider feature distribution and more balanced augmented image.By au gmenting t he f lag i mage i n the a ctual ap plication s cenario,the o ver-fitting problem of small sample images in the deep learning method is effectively alleviated,and the accuracy of the model is improved.