CN114549906A - Improved image classification algorithm for step-by-step training of Top-k loss function - Google Patents

Abstract

The invention discloses an image classification algorithm for step-by-step training of an improved Top-k loss function, which comprises an image data preprocessing module; the system comprises a deep learning feature extraction module and a system prediction output module, and is characterized in that an image data preprocessing module preprocesses input image data; the deep learning feature extraction module comprises an improvement of step-by-step training by using an improved Top-k loss function by using deep learning, and the classifier module is a deep learning network module; and the system output module processes the output of the classifier and outputs a judgment result. The image classification system breaks through the accuracy limit of the deep neural network under the condition of not modifying the network structure by using an improved image classification algorithm for step-by-step training of the Top-k loss function, and the accuracy is effectively improved by using the improved algorithm.

Description

An Image Classification Algorithm Based on Step-by-Step Training of Improved Top-k Loss Function

technical field

The invention belongs to the field of image classification, in particular to an image classification system based on a deep learning improved Top-k loss function for step-by-step training.

Background technique

In various fields of daily life, image classification plays an increasingly important role, and the technology of Image Classification emerges as the times require. To further achieve the correct classification of images, given a set of images each labeled as a single class, we make predictions on the class of a new set of test images, and measure the accuracy of the predictions The results become the focus of the research. Traditional image classification algorithms extract hand-designed features, which are limited and difficult to design, so they are still unable to perform some complex tasks, such as K-neighbor algorithm (KNN), support vector machine (SVM), etc. Algorithms and design are often difficult, and the combination of feature extraction and classifier algorithms is cumbersome, and it is difficult to achieve high classification accuracy.

In recent years, deep learning has been continuously applied to image classification systems. At present, deep learning methods have achieved many breakthroughs in practical applications, and have gradually become an important tool for artificial intelligence. Convolutional neural network is a kind of deep learning algorithm. Compared with traditional hand-designed features, deep features have the advantage of not requiring manual design of complex and time-consuming feature extraction algorithms, only need to design effective neural network models, and The classification accuracy is high. However, deep neural networks are often unable to break through the accuracy limit without modifying the network structure, and an improved algorithm is needed to effectively improve the accuracy.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a step-by-step training algorithm for an improved Top-k multi-loss function for image classification, which can achieve higher classification accuracy than the original loss function compared to the traditional method, and use the deep learning method at the same time. The steps of manual feature extraction and manual selection of classifiers are omitted, which overcomes the difficulty of feature extraction and classification in traditional methods.

The solution proposed by the present invention is to use an improved Top-k multi-loss function step-by-step training algorithm for image classification to achieve higher classification accuracy, including: image data preprocessing module; deep learning feature extraction module and system prediction output module.

The image data preprocessing module represents the image data as a three-dimensional tensor. The functions of the image data preprocessing module include: discarding irrelevant input data, which can reduce negative effects, usually by changing the brightness of the original image, changing the contrast of the original image, center trimming the image, changing the color function, denoising preprocessing. processing, etc. Preprocess the image data to convert it into an input form acceptable to the image target classification module. In the training mode, you need to label the image target data category and select the data set required by the machine learning method.

The working method of image data preprocessing in training mode is: for three-channel images, changing the brightness of the original image, changing the contrast of the original image, center trimming the image, changing the color function, denoising preprocessing, and uniform upsampling , center cropping, rotation and other preprocessing to expand the amount of data. In the training model, the parameters of the convolutional neural network can be updated to obtain a CNN network that can accurately classify the specified data set.

The working method of image target data preprocessing in test mode is: for three-channel images, change the brightness of the original image, change the contrast of the original image, change the function of color, denoise preprocessing, do unified upsampling, and do not do data expansion deal with. Uniform upsampling to a fixed resolution. In the following embodiments, the image resolution obtained after upsampling is 224*224 pixels. Get the final classification effect.

The image target classification module includes a deep learning feature extraction module and a loss function step-by-step training module, and a supervised learning algorithm is used to train the network. CNN network feature extraction module and classifier module, Top-k loss function is an improvement of cross entropy loss function, training a model with lower top-2 error is simpler than improving top-1 correct classification, and only the first two predictions are required. The category contains ground truth labels, which can effectively improve the fault tolerance of the model structure. The base classifier through a joint training strategy is used for our proposed ensemble learning, and multiple CNN models are trained separately using this training strategy. In the first step, we train a CNN network with a cross-entropy loss function initialized by an ImageNet pretrained model and fine-tuned for the dataset. We know that the cross-entropy loss function is a loss function created to optimize the top-1 loss, and the model is trained by the cross-entropy loss function to obtain a network with a higher top-1 correct rate. In the second step, we fine-tune using the top-k loss function created for the top-k correct labels, and use the model trained in the first step as the initial weights for the top-k loss function to train at the top-1 The optimized network is obtained with the same accuracy and higher top-2 accuracy, and the model improves the recognition ability. The network is jointly trained by two parallel loss functions to form a unified network.

The system output module processes the output of the classifier and outputs the judgment result.

Compared with the traditional cross-entropy loss function, an improved image classification algorithm based on the step-by-step training of the Top-k loss function disclosed in the present invention can achieve higher classification accuracy of Top-1 and Top-2, while reducing manual calculation. complexity. On the basis of changing the model architecture, the present invention improves the accuracy of Top-1 and Top-2 compared with the cross-entropy loss function. Compared with the traditional training method, the proposed improved Top-k loss function step-by-step training image classification algorithm has better robustness and better performance on the CIFAR-10 dataset.

Description of drawings

Figure 1 is a flowchart of an image classification algorithm using deep learning;

Fig. 2 uses two kinds of loss functions to carry out step-by-step training to realize the flow chart of image classification;

Figure 3 is a comparison chart of the accuracy of the traditional training method and the proposed Top-k loss function step-by-step training;

Detailed ways

The present invention will be described in further detail below with reference to the accompanying drawings.

First, the selection of the data set is carried out. The data set we choose is the CIFAR-10 image classification data set, which is a benchmark data set widely used in image classification. The dataset contains 60,000 images divided into 10 classes (airplane, car, bird, cat, deer, dog, frog, horse, boat, and truck), divided into 50,000 images for training and 10,000 images for testing Images, all 32x32 pixels in size. The dataset details are shown in the following table:

Table 1 CIFAR-10 dataset information table

data set Training set validation set test set category CIFAR-10 50000 0 10000 10

The choice of deep learning model, the deep learning model classifiers we selected for testing are:

Inception-v3, Inception-v3 has strong image feature extraction and classification performance, and is a widely used image recognition model. It consists of symmetric and asymmetric building blocks, including convolution, average pooling, max pooling, dropout, and fully connected layers. Batch normalization is widely used throughout the model and applied to the activation inputs. .

DPN92, the High Order RNN structure (HORNN) is used in the DPN model to connect DenseNet and ResNet, which proves that DenseNet can extract new features from the previous layer, and ResNet is essentially the feature that has been extracted in the previous layer. reuse. By combining the advantages of these two structures, the DPN network can effectively improve the classification efficiency.

ResNet, ResNet uses identity mapping to directly transmit the output of the previous layer to the back layer, which increases the depth of the network, and the error will not increase. The deeper network will not bring about an increase in the error on the training set, and solve the gradient disappear problem.

To train the model, use the environment training dataset shown in the table below.

Table 2 Experimental environment table

The model training parameters are shown in the following table. Adjust the Batchsize appropriately according to the model parameters to reduce the video memory usage of the graphics card and the model training time.

Table 3 Model initialization parameter table

parameter value learning rate 0.1 learning rate decay 0.1 momentum 0.9 batch size 32/24 training rounds 200

During training, an image classification algorithm trained in steps using our proposed improved Top-k loss function, as shown in Figure 2, training a model with lower top-2 errors is simpler than improving top-1 correct classification , and can effectively improve the fault tolerance of the model structure as long as the first two predicted classes contain ground truth labels. The base classifier through a joint training strategy is used for our proposed ensemble learning, and multiple CNN models are trained separately using this training strategy. In the first step, we train a CNN network with a cross-entropy loss function initialized by an ImageNet pretrained model and fine-tuned for the dataset. We know that the cross-entropy loss function is a loss function created to optimize the top-1 loss, and the model is trained by the cross-entropy loss function to obtain a network with a higher top-1 correct rate. In the second step, we fine-tune using the top-k loss function created for the top-k correct labels, and use the model trained in the first step as the initial weights for the top-k loss function to train at the top-1 The optimized network is obtained with the same accuracy and higher top-2 accuracy, and the model improves the recognition ability. However, due to the ambiguity of labels and the uncertainty of features, there is no way to really identify, except to improve the efficiency of feature extraction, the predicted labels are hidden between top-1 and top-k, and the improved top-2 classification accuracy is further To improve we will extract the top-2 labels for efficient use, the network is jointly trained by two parallel loss functions to form a unified network.

Test the model, after training the model, the test results of the three deep learning models of ResNet, DPN92, and Inception-v3 are obtained. The effect of using the traditional cross-entropy loss function is shown in the following table:

Table 3 Classification results on CIFAR-10 dataset using loss function

method Accuracy ResNet18 96.50 DPN92 97.56 Incption-v3 97.19

Test the model, after training the model, the test results of the three deep learning models of ResNet, DPN92, and Inception-v3 are obtained. The Top-k loss function proposed by us is used for step-by-step training as shown in the following table:

Table 4 compares the classification results using different loss functions on Inception-v3, DPN92 and ResNet18 models on the CIFAR-10 dataset.

This paper proposes an image classification algorithm with an improved Top-k loss function trained step by step. While ensuring the top-1 accuracy, it tries to reduce the top-2 generalization error of the model, and emphasizes the top-k loss by constraining the value. The importance of multiple output values of the model and the importance of the top-k loss further improve the classification performance. Contrastive experiments are conducted on the two loss functions on the CIFAR-10 dataset. The base classifier after using the top-k loss function is not only better than using the cross-entropy loss function in terms of top-1 accuracy and top-2 accuracy. It can be seen that the classification performance of the model is improved after applying the top-k loss function. In particular, after applying the top-k loss function, on the CIFAR-10 dataset, the top-1 accuracy of the model is improved by about 0.03%-0.2%, and the top-2 accuracy is improved by about 0.2%-0.6%, Fig. 3 shows the stability of the top-k loss function on the CIFAR-10 dataset.

Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, etc., All should be included within the protection scope of the present invention.

Claims (4)

1. the image classification algorithm of an improved Top-k loss function step-by-step training, is characterized in that, comprises module: (1) Image data preprocessing module (2) Deep learning feature extraction module (3) System prediction output module. 2. the image classification algorithm of a kind of improved Top-k loss function step-by-step training according to claim 1, is characterized in that, the image data preprocessing module in module (1), carries out preprocessing to the input image data , the deep learning feature extraction module in module (2), including the use of deep learning to use an improved Top-k loss function step-by-step training improvement, the classifier module is a deep learning network fully connected layer, module (3) The system prediction output module of the classifier outputs the judgment result after processing the output of the classifier. 3. the image classification algorithm of a kind of improved Top-k loss function step-by-step training according to claim 1, it is characterized in that by improving Top-k loss function step-by-step training can replace cross entropy loss function, constructs Top-k loss function step by step training. A better network with improved 1 accuracy and Top-2 accuracy. The traditional cross entropy loss function, the mathematical formula is expressed as:

The top-k loss depends on whether y is part of the top-k predictions, which is equivalent to comparing the top-k predictions with the ground-truth labels. The mathematical formula of the Top-k loss function is expressed as:

4. the image classification algorithm of a kind of improved Top-k loss function step-by-step training according to claim 1, is characterized in that can guarantee to improve the ability of model feature extraction while not revising network structure, obtain better. classification effect and improve the robustness of the model.

Images