Target Detection And Recognition Based On Deep Learning

With the advent of the artificial intelligence era,target detection and recognition in computer vision has increasingly become the focus of researchers.Target detection and recognition based on deep learning is far superior to traditional target detection and recognition methods in both detection accuracy and recognition speed.Although deep learning has been successfully applied to target detection and recognition tasks,there are still problems in detection accuracy and recognition speed that cannot be compared with humans.The performance of models and algorithms needs to be improved.Based on the target detection and recognition of region selection,this paper improves the convergence speed of deep learning detection model by improving the widely used stochastic gradient descent method in deep learning.Introducing new image enhancement methods and increasing the amount of model training image data is used to reduce the model overfitting phenomenon.The idea of small convolution is proposed to improve the detection accuracy of the detection model for small targets.The detailed research contents are as follows.Aiming at the ill-posed problems of stochastic gradient descent optimization algorithm in deep learning and the slow convergence of the saddle point in high-dimensional space,the exponential weighted moving average idea is introduced and a new optimization algorithm(Apro)is proposed.When the gradient parameters are updated,Apro not only considers the gradient effects at the current iteration but also the effects of the historical gradients.Apro gives different weights to the historical gradient,and the closer to the gradient of the current iteration,the greater the weight,which is used to solve the ill-posed problems in the stochastic gradient optimization algorithm.The model parameter update rule is modified to obtain the adaptive learning rate.Finally,the convergence theory of Apro algorithm is analyzed to prove the convergence of the Apro algorithm.The Apro optimization algorithm and the existing optimization algorithm are compared with the existing optimization algorithms on the MNIST dataset and CIFAR10 dataset to verify the performance of the optimization algorithm.Aiming at the problem of model over-fitting caused by insufficient training data in target detection and recognition,an image enhancement method based on Fourier transform is proposed.The Fourier transformed image is processed by a Gaussian filter to increase the amount of training set data and to attenuate the model overfitting problem.Introducing the idea of small convolution,three small convolution steps are added to the model of deep learning target detection to improve the detection accuracy of the model for small targets.Finally,the experiment is carried out on the target detection and identification task of sea cucumber.The model of this paper can achieve the target detection and identification of video level sea cucumber.Compared with the image selection model based on region selection,only the image data can be processed,and the detection speed of the model is greatly improved.