System simulation of computer image recognition technology application by using improved neural network algorithm

Digital image technology is penetrating into various fields of people's life, and it has been very mature and can effectively store and transmit data. Moreover, there are still various researches on image recognition, the core of this technology. The algorithm is mainly based on computer technology to obtain the target image for different scene categories, thus completely replacing the traditional classification form. Because of the limitations of traditional identification technology, there are some problems in the actual use. It does not depend on the prior knowledge requirements and can carry out complex feature space division. In this paper, an image recognition computer system is established by introducing an improved neural network algorithm. The algorithm is designed and tested, and the results show it has lower image recognition error rate. Subsequently, this research result is applied to the actual scene for testing. The test results show that the improved neural network optimization algorithm can make the extracted features more accurately expressed in the image processing, which is more effective than the traditional algorithm.


Introduction
Image recognition technology can be traced back to 1920s.Before 1950, due to the continuous improvement of computer technology, it has been applied in the eld of numerical calculation;In the 1960s, due to the emergence of new computer technology and the gradual development and perfection of theory, it was gradually applied to digital image processing [1].One of the main carriers of information transmission is image, which plays an important role in information interaction.In daily life, people will collect image information through various ways, such information can be applied to digital technology and image technology to meet people's information needs [2].Image recognition technology can effectively extract the features of the original image and classify based on the features [3].Before the formal classi cation, it is necessary to preprocess the original image and extract features, so as to create and optimize the automatic machine vision recognition system based on the above information.Image recognition covers many elds and can be applied to industrial process monitoring, weather forecasting, transportation, medical and military elds.Although there are abundant achievements in the eld of image recognition, there are also some problems [4].Parallel neural network can realize distributed storage and adaptive processing form for information, so it can realize processing and high error tolerance processing for massive data, and protect the image itself from damage and the information carried by the image from damage in this process.Therefore, the use of improved neural network technology algorithm in image recognition has great potential that other algorithms can not compare, and is an effective method to solve the problems related to image recognition.In short, using the improved neural network image recognition technology can effectively improve the e ciency of image recognition, has a wide range of applications.

Relevant Work
The literature continuously deepens the structural layers of the model to improve its performance, and increases the number of parameters in the process.After the literature reading and research, the basic research model is MobileNet.This model can replace the traditional convolution calculation process, so that it is deeper and separable, and can combine two-dimensional and three-dimensional convolution operations [5].Since the traditional convolution operation is a combination of addition and multiplication, the application of the new algorithm can greatly reduce the model parameters and the number of calculations, while maintaining the exact requirements of the image recognition task.On the basis of the original model, this paper improves the clustering middle layer model and tries to optimize the related problems [6].The literature has conducted in-depth research on various popular deep neural network optimizers, summarized and analyzed the advantages and disadvantages of each optimizer, it is proposed that RMSProp optimizer (in the original MobileNet) can be used to improve the comprehensive performance of the current algorithm [7].After using the optimizer, the experimental results show that the training stability and target accuracy of the algorithm model are effectively improved.Based on the analysis of traditional convolutional neural network, this paper uses MobileNet to optimize the neural network model, so as to complete the model recognition [8].After analyzing the neural network optimization algorithm and MobileNet, an improved algorithm for MobileNet is further proposed.Through experiments, it can be seen that Adam optimizer has very good comprehensive performance, and can improve the accuracy and convergence speed of the model by combining with the migration learning algorithm.Design experiments, comparing the optimization algorithm with the traditional MobileNet model, we can see that the improved algorithm can better complete the task and improve the recognition accuracy, and there are many improvements in all aspects [9].The literature describes the characteristics of different detection algorithms,a convolutional neural network is proposed to optimize the algorithm to complete the construction of tensor ow network system,which completes the programming in pycharm environment and the hardware implementation based on Xavier module [10].The literature introduces the SVM generalization error strategy and applies it to the learning process of the CNN depth model, so as to guide the construction process of the CNN depth model and related classi ers.Then, a class of CNN model is proposed, which is driven by RMB.Compared with the traditional model, it can better limit the expansion of the radius of the MEB, and increase the classi cation range between different types of image features, and nally promote CNN to obtain higher quality image features [11].The literature uses minimum class classi cation variance SVM combined with Fisher linear discriminant theory to guide the deep learning of CNN, and proposes a CNN model combined with mcvsvm model.This model can not only further consider the range of classi cation training, but also adjust the hyperplane according to the distribution of samples, so as to obtain better hyperplane classi cation to guide the deep learning of CNN model.Finally, CNN is used to extract more distinctive image features [12].In this paper, the integration idea is applied to improve the traditional convolutional neural network algorithm, and then the system design is completed based on this algorithm based on vgg-16 and improves the downsampling process to study the pooling algorithm, and obtains the downsampling characteristic map by assigning weights to each pooling element.It can overcome the key feature missing problem caused by the maximum pooling algorithm and the large eigenvalue problem caused by the weakening average pooling algorithm.
The improved algorithm is applied to the recognition of cervical cancer images [13].The experimental results show that the recognition rate of the algorithm is greatly improved after the introduction of convolutional neural network technology,which can reach 84.61%.The framework of residual neural network is established in the literature.The gradient descent algorithm is analyzed and improved, and an improved algorithm based on residual neural network is proposed.The in uence of noise can be eliminated based on the previous gradient; Every time the parameters are updated, several groups of samples are randomly selected for iteration to prevent the solution from falling into the local optimal situation [14].The improved algorithm is introduced in the recognition process of cervical cancer target image.The experimental results show that the optimization algorithm can improve the recognition accuracy of target image, effectively improve the recognition e ciency and reduce the false recognition rate.The detection accuracy of the algorithm is 96.83% 3. Basic Model And Improvement Of Neural Network For Image Task

Convolutional neural network training method
In the deep learning theory, the network model constructs neurons layer by layer.Each time a layer is trained, the lower layer will receive the error from the upper layer and adjust it.In the process of network training, it mainly includes forward propagation and reverse propagation.The former mainly extracts the target features and classi es the results, while the latter mainly modi es the model parameters.
The loss function will avoid over tting by adding a regular term or a weight loss term, as shown in formula 4: 1 l and i, j denote the layer index and neuron index, respectively, and the correlation between the regularization term and the square difference term is determined by the weight decay parameter λ Control.
The adjustment direction of W: For each layer of neurons, de ne δ:  Obviously, the standard SVM model is transformed from the original complex optimization problem to a simple convex quadratic programming problem.After problem ( 11) is solved, a small Lagrange multiplier vector is obtained, in which a small number of Lagrange multipliers are not 0, so as to obtain the key support points of the vector in the feature space, and nally establish the decision management program, which is equal to the corresponding compensation period in the plan:

Support vector machine driven convolutional neural network
12 In most deep convolution models, the combination of Softmax and cross entropy is often used to guide CNN training.Therefore, the following objective functions are de ned: 13 Analyzing the experimental results and describing the in uence of iteration times on the recognition results of the algorithm itself, it can be seen that the recognition accuracy of the algorithm is also improving with the increase of iteration times.
The identi cation error rate under each iteration number is shown in Table 1:   1 and Table 2 that the increase of iteration times will lead to the decrease of the detection error rate of the same algorithm, and the decrease percentage of the error rate gradually tends to 0. It can be seen that the stability of the model also gradually becomes stronger.

Image Recognition System By Using Deep Neural Network And Its
4.1 Image recognition system architecture on the above work, this paper trains a computer image recognition technology system that supports 22000 kinds of object recognition.The system framework is shown in Fig. 2 below.

Basic process of image recognition
short, through image recognition, feature recognition and classi cation can be carried out for the research object, such as recognition of common objects in life: handwritten digits.With the rapid development of related recognition technology, the types and speci c requirements of machine recognition are increasing, resulting in more complex recognition process.At present, image conversion or data extraction for speci c objects is the main task of machine recognition, that is, image and digital processing.The ultimate goal is to obtain features and classify them.
Figure 3 shows the image recognition process.
The purpose of image preprocessing is to denoise the target image to reduce the complexity and redundant data of subsequent operations and improve the recognition accuracy and performance of the algorithm.During the use of image recognition algorithms, redundant information that interferes with the motion the algorithm will appear.It is necessary to process this part of the content to input the model, so as to highlight the target features to be extracted.Feature extraction is the most important step and the most important goal in image recognition algorithm, which will affect the nal output result of the model.Feature extraction can transform the original image information into information that can be recognized by the computer.This process can ensure a large Euclidean distance between the features between the classes and the acquired features, thus enhancing the difference.Image features are divided into two categories: global features and local features.The task of image classi cation is to classify the obtained target features after feature extraction.This process needs to train the classi er for the training data set, and apply it to the classi cation task after the effective parameters are obtained after the training, so as to complete the classi cation.
The most commonly used image sharpness processing method is Laplace ltering, which belongs to linear spatial ltering.The Laplace transform of the image function f (x, y) can be expressed as:

Simulation experiment results and analysis
As shown Table 3.The growth curve of the model recognition rate is shown in Fig. 5: The results of Algorithm classi cation ability are shown in Table 4.The accuracy of the improvement model is about 1.3%higher.It is higher than the Alexnet model when the parameter memory consumption is greatly reduced and the training iteration time is shortened by nearly 90%.The test results are shown in Table 5.

Image recognition technology based on networks
Traditional -based is mainly to retrieve the most related images from query keywords from the image database.The query task is considered a very di cult and complex and abstract task.One possible method is to use the natural language description of the network image environment as an image description, and the image extraction task is performed according to the method of text extraction.
This method of image capture is greatly impacted by the quality of the webpage, and it is easy to insert irrelevant images.In order to improve the performance of image extraction and reduce the existence of unrelated images in the extraction result, this article proposes a variety of classi cation algorithms based on low -level image features.
When using deep neural network processing images, we can determine the decomposition relationship from part to overall.Based on this feature, this topic proposes a new DNN training method with multitasking architecture to help complete image collection tasks.The current commonly used DNN architecture is the form of a multi -classi er, that is, the last layer of the neural network corresponds to multiple categories and the probability of the corresponding prediction category.Due to the large number of DNN parameters, the number of forecast categories generally supported by the current single GPU system is limited.Due to the large number of queries in practical applications, this architecture is not suitable for a large query task.In addition, real data categories are different from the query process, but are similar to visual features, such as "dog" and "puppy".For these categories, the DNN of the multiclassi er can be used directly in the training process.
Each object of Multi -task DNN corresponds to a dual -class task (based on the original training sample, remove the negative cases of other categories, calculate the correlation between the respective training and calculation images and categories), and consider the characteristics of the between many objects can be shared.For example, the tire characteristics recognized in motorcycle categories are also applicable to vehicle recognition.We have proposed a new training method, Ring-Training, trying to continuously transfer and adjust behavior between different tasks.In the process, some stronger general features can be learned.

Conclusion
The traditional method of manual feature extraction is not only cumbersome and error prone, but also wastes a lot of time, energy and resources.The accurate recognition of target images has important practical signi cance for various research elds.The improved convolutional neural network image recognition algorithm gradually shows its advantages and is widely used in various industries.The improved neural network optimization algorithm is used to process images so that the extracted features can be more accurately expressed.Aiming at the computer image recognition model, this paper makes an in-depth study, and then combines the improved neural network algorithm with the image recognition technology to complete the model design.The experiment proves that this kind of model has a good recognition effect.Growth curve of model recognition rate

4 δ
j is calculated by chain rule: Where the change direction of W, b is: the BP algorithm repeatedly uses the chain rule to calculate the cumulative gradient loss and nally minimize the loss function.Input the reference set into the non-optimized neural network, and conduct training, learning and feature extraction on different scenes from the data set.After several iterations, you can see the loss value obtained, x the model hyperparameters, change the learning rate to 0.001, and change the time value to 30, and get the accuracy rate curve and loss curve in the training stage, as shown in Fig. 1.
Find the partial derivative of formula(10) with respect to W and b, and make it 0 to obtain: 10 Substitute equation and obtain the dual optimization problem of the standard SVM optimization model after simpli cation: 11

3Figure 4
Figure4shows the simulation results of the SVHN street view digital data set.The size of the image block is {50,100,200,500,1000}, and compared with the CNN -driven training process in SVM.Among them, the SVM -driven CNN training block size is set to 500.

Figures Figure 1 vggnet accuracy and loss curve Figure 2 Architecture of computer image recognition technology system Figure 3 Flow
Figures

Table 1
Recognition error rate of algorithm under different iteration times

Table 5
Under actual environmental conditions, the recognition of Alex Net models and improvement of neural network models