The invention belongs to the computer image processing technology field and particularly relates to an image grain size classification method based on the jumpable double-branch neural network. The method comprises steps that a data set is collected, and image data and corresponding coarse and fine category labels are included; the jumpable double-branch neural network is constructed, and a hierarchical gate control module is added to the densely-connected convolutional neural network; data enhancement of a training sample is performed, and the jumpable double-branch neural network is trained;an end-to-end neural network is trained, and the classification result is outputted. Coarse-grained and fine-grained classification tasks of images can be simultaneously solved through the constructed jumpable double-branch neural network. The method is advantaged in that the gate control module is added to the densely-connected convolutional layer, execution of the convolutional layer can be dynamically selected based on the input, the inference depth of the network is variable depending on complexity of the sample and the granularity of the classification tasks, and the flexible convolutional neural network is beneficial for improving model capacity and computational efficiency.

Image grain size classification method based on jumpable double-branch neural network

The invention relates to a target detection model based on deep learning and a training method thereof. According to the model, the structural advantages of dense connection are fully utilized, the training starts from zero, and an end-to-end detection effect can be achieved. According to the model, in the back-end feature extraction, a dense connection module is used to establish more associations between adjacent convolutional layers, model performance is improved and weight parameters are reduced. In front-end feature map merge, a dense connection mode is used to carry out feature reuse, the feature of a previous feature map is introduced into the feature map of each scale, a connection mode with a shorter path with a top layer is established, and thus an error signal is transmitted deeper along the network in the reverse propagation. The model has fewer parameters and stronger performance, and the disadvantage that a traditional model relies heavily on pre-training is made up. Themodel and the training method are also applicable to a target detection task of special domain images, and the problem of no convergence or serious over-fitting in training starting from zero is improved.

Target detection model based on deep learning and training method thereof

The invention provides an image classification model design method and apparatus The method comprises the steps of according to a labeled image database, performing training to obtain an initial network model of image classification; according to a prediction result of the initial network model and a prediction result of a final network model, building a loss function corresponding to the final network model; and according to the loss function, optimizing the final network model According to the image classification model design method and apparatus provided by the invention, the requirementon a training sample quantity is relatively low; the design difficulty and cost are lowered; and the problem of labeling ambiguity is effectively solved

Image classification model design method and apparatus

The invention relates to a Mobile Nets-based method for optimizing a multi-scale learning network. According to the invention, the multi-scale learning network comprises four parts. The front three parts of the four parts are composed of identical separable convolution layers, wherein the rear par of each separable convolution layer is connected with bachnorm and ReLU, then connected with a pooling layer, and finally connected with the full-connection layer and the output layer of the fourth part. The separable convolution layer comprises three groups of convolution operations. The specific network structures of the three groups of convolution operations are as follows: the first group is subjected to convolution operation according to the depth convolution of 3 * 3; the second group is subjected to convolution operation according to two continuous depth convolutions of 3 * 3; the outputs of the first and second groups are first subjected to adding operation and then subjected to convolution operation according to the point convolution of 1*1; the third group is subjected to convolution operation directly according to the point convolution of 1*1; the outputs of the first, second and third groups are subjected to combining operation. The experiment comparison shows that, the network structure constructed by the method and disclosed by the invention is few in experimental parameters and high in precision. The three groups of separable convolution layers are stable in structure and the experiment effect is most ideal.

Mobile Nets-based method for optimizing multi-scale learning network

The invention discloses a rapid target detection method based on a lightweight neural network. The method comprises the steps of constructing a convolutional neural network, training the convolutionalneural network and performing target detection; defining main modules of Front module and Tinier module of the convolutional neural network; defining the number of layers of the convolutional neuralnetwork and a pooling layer position; preprocessing a data set, and converting the data set into a standard format input by a target detection framework Darknet; initializing a parameter of the convolutional neural network; training a neural network model through continuous iteration of forward propagation and back propagation; inputting a test image, and calculating by use of the neural network model obtained in a training process to obtain a detection numerical result; and making annotation on the image according to the detection result, and marking a position and a category of each object by use of a rectangular box. By use of the method, the detection speed is faster and the accuracy rate is higher as well.

Rapid target detection method based on lightweight neural network

The invention discloses a quick image classification method based on deep learning, and the method comprises the following steps: 1, network building; 2, data set preprocessing; 3, network training; 4, image classification, wherein the image classification comprises the following substeps: 4.1, inputting a testing data set into the trained network model after preprocessing, and extracting the multi-scale features of a testing image; 4.2, inputting the extracted multi-scale features into a Softmax classifier, and outputting the probability that the testing image belongs to one class; 4.3, inputting probability that the testing image belongs to one class and a label corresponding to the image into an Accuracy network layer, and outputting the probability that the image is classified correctly. Through the above steps, the method can achieve the quick classification of the testing image. The method is short in training time, is convenient in training, and is high in classification precision.

Quick image classification method based on deep learning

The invention relates to a method for optimizing a deep convolutional neural network for image classification, which comprises the steps of first, building an image classification convolutional neural network; second, training the image classification convolutional neural network; third, testing the image classification convolutional neural network. The process of testing the image classification convolutional neural network comprises the steps that a preprocessed test data set is fed into a trained network model, and an Accuracy layer of the network outputs an accuracy value according to a probability value outputted by a Softmax layer and a label value of an input layer, wherein the accuracy value is the probability that a test image is correctly classified. According to operations performed in the steps, optimization for the deep convolutional neural network for image classification can be realized. The method provided by the invention for optimizing the deep convolutional neural network for image classification can effectively reduce the semantic gap and is high in classification accuracy.

Method for optimizing deep convolutional neural network for image classification

The invention discloses a content-based image retrieval method based on significance segmentation The method comprises a training process and a testing process The training process comprises the following steps that firstly, a visual vocabulary dictionary is established, specifically, all images in a training image set are subjected to analysis and processing sequentially, and the visual vocabulary dictionary is established for subsequent retrieval; secondly, target foreground images and background regional images are obtained through segmentation by means of visual significance features of the images; thirdly, color features and textural features are extracted from the target foreground images and the background regional images; and fourthly, statistics of visual vocabulary distribution histograms of all the images in the database are made on the basis of operation in the first step The testing process comprises the fifth step that on the basis of operation in the four steps, tested images are retrieved By adoption of the content-based image retrieval method based on significance segmentation, the number of semantic gaps can be effectively decreased, and the accuracy is high

Content-based image retrieval method based on significance segmentation

The invention discloses a content-based image retrieval method for unsupervised adversarial training. The method comprises the following steps of: 1, carrying out network construction, wherein an unsupervised adversarial network framework consists of a generation model and a discrimination model, and both the generation model and the discrimination model are formed by three-layer full-connection networks; 2, carrying out data preprocessing; 3, carrying out network training: 3.1, initializing generation model and discrimination model parameters by using random weight values, 3.2, training the generation model, and 3.3, training the discrimination model; and 4, carrying out precision test. The content-based image retrieval method for unsupervised adversarial training is relatively good in robustness and relatively low in training data requirement, and does not need a large amount of label information.

Content-based image retrieval method for unsupervised adversarial training

Provided is an image retrieval method based on visual significant integration. A retrieval object is color images. The image retrieval method comprises a training process and a testing process. The training process comprises step 1, building a visual word dictionary: orderly analyzing all images in a training image set so as to set up the visual word dictionary for subsequent retrieval; step 2, utilizing a fusion algorithm of image visual saliency characteristics to obtain a visual saliency picture and obtaining a foreground target picture and a background area picture by partitioning of the saliency picture; step 3, respectively extracting image color characteristics and image texture characteristics of the foreground target picture and the background area picture; step 4, counting a visual word distribution histogram of each picture in a database based on operation of step 1. The testing process also comprises the step 5 of retrieving testing images. The invention provides the image retrieval method based on visual significant integration which is good in partitioning effect and high in retrieval performance.

Hao Pengyi

Papers

Quick image classification method based on deep learning

Method for optimizing deep convolutional neural network for image classification

Content-based image retrieval method based on significance segmentation

Content-based image retrieval method for unsupervised adversarial training

Image retrieval method based on visual saliency fusion