CN105894012A - Object identification method based on cascade micro neural network - Google Patents

Abstract

The invention relates to an object recognition method based on a cascaded microneural network. The steps are as follows: calculate the number of cascaded convolution layers in a deep neural network according to the size of an input sample image; construct a deep cascaded microneural network; set the depth The training parameters of the cascaded micro-neural network are trained using the stochastic gradient descent method; the softmax classifier is used for classification, and the forward propagation algorithm is used to calculate the classification error; the weight value of the parameters that need to be trained in the neural network is updated by using the back propagation operation ; A deep cascaded microneural network is obtained, which can efficiently identify objects of the corresponding category. The invention can improve the recognition performance of the object recognition system without increasing the parameters (calculation complexity). At the same time, the method is simple, easy to converge, and will not increase the training time.

Description

Object Recognition Method Based on Cascaded Micro Neural Network

Technical field

The invention relates to a fast and efficient object recognition method in the fields of human-computer interaction, computer vision, etc., in particular to a method for object recognition using a deep neural network.

Background technique

Object recognition is a crucial research field in computer vision, including license plate recognition, road sign recognition, biological species recognition, and multi-type object recognition in natural scenes. Object recognition can be widely used in security monitoring, intelligent transportation, human-computer interaction and other fields. For example, the rapid development of unmanned vehicles requires object recognition technology, which can help unmanned vehicles to identify objects on the road, so as to ensure that they can drive smoothly and safely.

At present, deep neural network technology has made breakthroughs in the field of object recognition. In 2012, Alex Drizhevsky et al. [1] designed a 7-layer neural network with 60 million parameters. The network can recognize 1000 different objects and has good recognition performance. In 2013, Network inNetwork (NIN) [2] used a multi-layer perceptron (MLP) to strengthen the association between different feature channels of the neural network, so that the neural network can learn more robust features and improve recognition performance. In 2014, the VGG neural network [3] extended the depth of the neural network to 19 layers, which greatly improved the performance of the object recognition system. In 2014, GoogleNet[4] constructed a 22-layer deep neural network using multi-scale convolution kernels, and won the object recognition championship in the ImageNet Large-Scale Visual Recognition Challenge 2014 (ILSVRC2014) competition. In 2015, ResNet [5] built a neural network with a depth of up to 152 layers, and won the object recognition championship in the ILSVRC2015 competition.

From the above related studies, it can be seen that increasing the depth of the neural network can improve the performance of the object recognition system. However, increasing the depth of the neural network also brings two problems. On the one hand, the parameters (calculation amount) of the neural network will be greatly increased, which will reduce the real-time performance of the neural network and cannot be effectively used in the real environment. On the other hand, the increase of neural network parameters makes it difficult for the neural network to converge and not easy to train, which will consume a lot of energy in the training process.

The existing neural network improves its recognition performance by increasing its depth. The fundamental reason is that the expressive ability of the basic operation unit (filter) is limited. Given an input image (or a video frame), the neural network needs to extract features from the image. Feature extraction is based on the local image block as the basic processing unit. The size of the filter of the existing neural network is consistent with the size of the local image block, and the output result is a feature value after a simple dot product operation. This feature transformation is very simple and has limited expressive power.

references:

[1] A. Krizhevsky, I. Sutskever, and G. Hinton, Imagenet classification with deep convolutional neural networks [C]. In Proceedings of Advances in Neural Information Processing Systems, 2012.

[2] M.Lin, Q.Chen, and S.Yan, Network in network[C], In Proceedings of International Conference on Learning Representations, 2014.

[3]K.Simonyan andA.Zisserman,Verydeep convolutional networks for large-scale image recognition[C],In Proceedings ofInternational Conference on Learning Representations,2015.

[4] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich, Going deeper with convolutions [C], In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015.

[5] K.He, X.Zhang, S.Ren, J.Sun, Deep Residual Learning for Image Recognition, arXiv:1512.03385, 2015.

Contents of the invention

The object of the present invention is to overcome the above-mentioned shortcomings of the prior art and provide a method that can improve the recognition performance of an object recognition system without increasing the complexity of computation. Technical scheme of the present invention is as follows:

An object recognition method based on a cascaded micro-neural network, the method adopts a deep neural network for object recognition, and its basic operation unit is composed of a cascaded sub-block filter f _l , and the i-th in the cascaded sub-block filter f _l sub-block filters w _i,k are composed of two basic filters and composed of k indicates the feature channel index of the output, It is called a spatial filter, its size is h _i ×w _i , and it satisfies h _i <H and w _i <W, h _i and w _i are the height and width of the spatial filter, respectively, and H and W are the input partial images The height and width of the block are used to extract the spatial characteristics of the sub-blocks in the input local image block; It is called a channel filter, and its size is 1×1, which is used to strengthen the connection between the output channel features. The size of the i-th sub-block filter w _i,k is expressed as (h _i ×w _i ,1×1), and the level The joint sub-block filter f _l is composed of n sub-block filter cascades, expressed as f _l =[(h ₁ ×w ₁ ,1×1),(h ₂ ×w ₂ ,1×1),... (h _n ×w _n ,1×1)], l represents the number of layers of the cascaded sub-block filter in the deep neural network; the cascaded sub-block filter f _l and the local image block of size H×W satisfy and The constraints of the cascaded sub-block filter f _l and input channel feature convolution to obtain a cascaded convolutional layer, denoted as C _l ;

The steps of the object recognition method are as follows:

Step 1: According to the size H _I and W _I of the input sample image, H _I represents the height, and W _I represents the width, calculate the number L of cascaded convolutional layers in the deep neural network, where L satisfies 2 ^L × 5 ≥ H _I , the minimum integer value of 2 ^L ×5≥W _I ;

Step 2: Construct a cascaded filter f _l according to the size H×W of the local image block processed by each cascaded convolutional layer C _l , and perform a 0.5 times downsampling operation on the obtained cascaded convolutional layer C _l , Connect all L cascaded convolutional layers C _l to construct a deep cascaded microneural network;

Step 3: Set the training parameters of the deep cascaded micro-neural network, and use the stochastic gradient descent method for training;

Step 4: Use the softmax classifier for classification, and use the forward propagation algorithm to calculate the classification error;

Step 5: Utilize the backpropagation operation to update the weight values of the parameters to be trained in the neural network;

Step 6: Repeat steps 4 to 5 until the verification error no longer changes, at this point the training process ends, and a deep cascaded microneural network is obtained, which can efficiently identify objects of the corresponding category.

By adopting the method of the invention, the traditional convolution operation is replaced by the cascaded micro-neural network operation unit, and the recognition performance of the object recognition system can be improved without increasing parameters (complexity of operation). At the same time, the method is simple, easy to converge, and will not increase the training time.

Description of drawings

Fig. 1 is the block diagram of proposed method of the present invention

detailed description

The present invention will be described below in conjunction with the accompanying drawings and embodiments.

The cascaded micro-neural network proposed by the present invention is different from the existing neural network, and its main difference is that the present invention uses cascaded sub-block filters instead of traditional filters. That is, the basic operation unit of the cascaded micro-neural network proposed by the present invention is composed of cascaded sub-block filters. For the convenience of later description, some concepts and variables are briefly explained first. The i-th sub-block filter w _i,k in the cascade of sub-block filters consists of two basic filters and composition, which can be expressed as k indicates the feature channel index of the output. It is called a spatial filter, its size is h _i ×w _i , and it satisfies h _i <H and w _i <W. h _i and w _i are the height and width of the filter, respectively, and H and W are the height and width of the input local image block. Its role is to extract the spatial features of the sub-blocks in the input local image patch. It is called a channel filter with a size of 1×1, and its function is to strengthen the connection between the output channel features. The size of the sub-block filter w _i,k can be expressed as (h _i ×w _i , 1×1). The cascaded sub-block filter is composed of n sub-block filters cascaded, which can be expressed as f _l =[(h ₁ ×w ₁ ,1×1),(h ₂ ×w ₂ ,1×1),... (h _n ×w _n ,1×1)], l represents the number of layers of the cascaded sub-block filter in the deep neural network. The cascaded sub-block filter f _l and the local image block of size H×W satisfy and constraints. The cascaded sub-block filter f _l is convolved with the input channel features to obtain a cascaded convolutional layer, denoted as C _l .

The present invention replaces traditional filters with cascaded sub-block filters, and can extract more robust local features without increasing system parameters. The specific steps are as follows:

Step 1: Calculate the number L of cascaded convolutional layers in the deep neural network according to the dimensions H _I and W _I of the input sample image (H _I represents the height, and W _I represents the width). Where L is the smallest integer value that satisfies 2 ^L ×5≥H _I , 2 ^L ×5≥W _I.

Step 2: Construct a cascaded filter f _l according to the size H×W of the local image block processed by each cascaded convolutional layer C _l , and perform a 0.5 times downsampling operation on the obtained cascaded convolutional layer C _l . Connect all L cascaded convolutional layers C _l to build a deep cascaded microneural network.

Step 3: Set the training parameters of the deep cascaded micro-neural network. We use stochastic gradient descent for training. The number of training samples in each batch is m=100, the training momentum is 0.9, and the weight decay factor is 0.0005. We initialize the weight values of all training parameters using a Gaussian distribution with mean 0 and standard deviation 0.01. The initial learning rate is 0.01.

Step 4: Use the softmax classifier for classification, and use the forward propagation algorithm to calculate the classification error.

Step 5: Utilize the backpropagation operation to update the weight values of the parameters to be trained in the neural network.

Step 6: Repeat steps 4 to 5 until the verification error no longer changes, at this time the training process is over, and a deep cascaded microneural network can be obtained. The cascaded microneural network can efficiently identify objects of the corresponding category.

Claims (1)

1. an object identification method based on the micro-neutral net of cascade, the method uses deep neural network to carry out object identification, its elementary operation list Unit is by cascade sub-block wave filter f_lComposition, cascades sub-block wave filter f_lIn i-th sub-block wave filter w_i,kBy two primary filtersWith Composition, is expressed asThe feature passage index of k instruction output,Being referred to as spatial filter, size is h_i×w_i, and meet h_i＜ H and w_i＜ W, h_iAnd w_iBeing respectively height and the width of this spatial filter, H and W is height and the width of topography's block of input, For extracting the space characteristics of sub-block in input topography block；Being referred to as path filter, size is 1 × 1, is used for strengthening output channel feature Between contact, i-th sub-block wave filter w_i,kSize be expressed as (h_i×w_i, 1 × 1), cascade sub-block wave filter f_lBy n sub-blocking filter cascade Composition, is expressed as f_l=[(h₁×w₁,1×1),(h₂×w₂,1×1),...(h_n×w_n, 1 × 1))], l represents that cascade sub-block wave filter is residing in deep neural network The number of plies；Cascade sub-block wave filter f_lMeet with topography's block that size is H × WWithConstraint, cascade sub-block wave filter f_lObtain concatenated convolutional layer with input channel feature convolution, be expressed as C_l。

The step of described object identification method is as follows:

Step 1: according to size H of input sample image_IAnd W_I, H_IRepresent height, W_IRepresent width, calculate deep neural network cascade The number L of convolutional layer, wherein L is for meeting 2^L×5≥H_I,2^L×5≥W_ISmallest positive integral value；

Step 2: according to each concatenated convolutional layer C_lSize H × the W of handled topography's block builds cascading filter f_l, to the cascade obtained Convolutional layer C_lCarry out the down-sampled operation of 0.5 times, connect all L concatenated convolutional layer C_lBuild a degree of depth and cascade micro-neutral net；

Step 3: set this degree of depth and cascade the training parameter of micro-neutral net, uses stochastic gradient descent method to be trained；

Step 4: use softmax grader to classify, and utilize propagated forward algorithm to calculate error in classification；

Step 5: utilize back propagation computing to update the weighted value of the parameter needing training in neutral net；

Step 6: repetition step 4 is to step 5, till validation error no longer changes, now training process terminates, and obtains a degree of depth cascade Micro-neutral net, the micro-neutral net of this cascade can identify the object of respective classes efficiently.