阅读说明：本技术 一种基于深度学习的藏医尿液颜色自动识别方法 (A kind of Tibetan medicine's urine color automatic recognition method based on deep learning ) 是由 刘勇国 张艺 杨尚明 李巧勤 华尔江 泽翁拥忠 降拥四郎 杜春慧 朱嘉静 郑子强 于 2019-10-14 设计创作，主要内容包括：本发明提供了一种基于深度学习的藏医尿液颜色自动识别方法,包括如下步骤：采集尿液图像,并对每幅图像进行标注；根据所述尿液图像构建基于深度卷积神经网络CNN的尿液颜色识别模型；根据所述尿液颜色识别模型识别尿液颜色特征,从而完成基于深度学习的藏医尿液颜色的自动识别。本发明通过以上设计能自动识别尿液在尿热阶段的颜色,通过深度卷积神经网络从训练数据中自动学习特征,而不是采用手工设计的特征,学习到的特征对于尿液颜色识别更有效,为医师诊断提供辅助支持。此外,所提出的算法直接在原始像素上运行,不需要预处理技术,在一定程度上节约时间。(The present invention provides a kind of Tibetan medicine's urine color automatic recognition method based on deep learning includes the following steps: to acquire urine image, and is labeled to each image；Urine color identification model according to the urine picture construction based on depth convolutional neural networks CNN；Urine color characteristic is identified according to the urine color identification model, to complete the automatic identification of Tibetan medicine's urine color based on deep learning.The present invention is by design energy automatic identification urine above in the color for urinating the hot stage, by depth convolutional neural networks from training data automatic learning characteristic, rather than the feature of hand-designed is used, the feature learnt identifies more effectively urine color, provides auxiliary for doctor diagnosed and supports.In addition, the algorithm proposed is directly run on original pixels, preconditioning technique is not needed, saves the time to a certain extent.)

1. a kind of Tibetan medicine's urine color automatic recognition method based on deep learning, which comprises the steps of:

S1, acquisition urine image, and each image is labeled；

S2, the urine color identification model according to the urine picture construction based on depth convolutional neural networks CNN；

S3, urine color characteristic is identified according to the urine color identification model, to complete Tibetan medicine's urine based on deep learning The automatic identification of liquid color.

2. Tibetan medicine's urine color automatic recognition method according to claim 1 based on deep learning, which is characterized in that institute It states and each image is labeled in step S1, specifically:

It is yellow, white, red, redness of the skin or complexion, brown, black, green, cyan and blue figure by urine image labeling collected Picture, and by the size of all urine Image Adjustings to 224 × 224.

3. Tibetan medicine's urine color automatic recognition method according to claim 1 based on deep learning, which is characterized in that institute Stating the urine color identification model constructed in step S2 includes sequentially connected input layer, the first convolution unit C1, the second convolution Unit C2, third convolution unit C3, Volume Four product unit C4, the 5th convolution unit C5, the first full connection unit FC1, second are entirely The full connection unit FC3 of connection unit FC2, third and output layer.

4. Tibetan medicine's urine color automatic recognition method according to claim 3 based on deep learning, which is characterized in that institute Stating the first convolution unit C1 includes sequentially connected first convolutional layer Conv1, the first non-linear layer Relu1 and the first pond layer Pool1, the first convolutional layer Conv1 include that size is 11 × 11,96 convolution kernels that step-length is 4, first pond layer The step-length of Pool1 is 2；

The second convolution unit C2 includes sequentially connected second convolutional layer Conv2, the second non-linear layer Relu2 and second Pond layer Pool2, the second convolutional layer Conv2 include that size is 5 × 5,256 convolution kernels that step-length is 1, second pond The step-length for changing layer Pool2 is 2；

The third convolution unit C3 includes sequentially connected third convolutional layer Conv3 and third non-linear layer Relu3, and described Three convolutional layer Conv3 include that size is 3 × 3,384 convolution kernels that step-length is 1；

The Volume Four product unit C4 includes sequentially connected Volume Four lamination Conv4 and the 4th non-linear layer Relu4, and described Four convolutional layer Conv4 include that size is 3 × 3,384 convolution kernels that step-length is 1；

The 5th convolution unit C5 includes sequentially connected 5th convolutional layer Conv5, the 5th non-linear layer Relu5 and third Pond layer Pool3, the 5th convolutional layer Conv5 include that size is 3 × 3,256 convolution kernels that step-length is 1, the third pond The step-length for changing layer Pool3 is 2；

The first full connection unit FC1 includes the sequentially connected first full articulamentum fc1 and the 6th non-linear layer Relu6；

The second full connection unit FC2 includes the sequentially connected second full articulamentum fc2 and the 7th non-linear layer Relu7；

The full connection unit FC3 of third includes the full articulamentum fc3 of third connecting with the output layer.

5. Tibetan medicine's urine color automatic recognition method according to claim 4 based on deep learning, which is characterized in that institute State the first convolutional layer Conv1, the second convolutional layer Conv2, third convolutional layer Conv3, Volume Four lamination Conv4 and the 5th convolution The output mapping of layer Conv5Are as follows:

Wherein, M^l-1Indicate the feature quantity of the urine color of input, l and m respectively indicate the index of the number of plies and urine characteristic pattern, n Indicate the index of filter,Indicate that filter, * indicate convolution operation,Indicate m layers of first of urine input vector,Indicate the biasing of l layers of n-th of urine output feature.

6. Tibetan medicine's urine color automatic recognition method according to claim 4 based on deep learning, which is characterized in that institute State the first non-linear layer Relu1, the second non-linear layer Relu2, third non-linear layer Relu3, the 4th non-linear layer Relu4, The activation primitive f (y) of five non-linear layer Relu5, the 6th non-linear layer Relu6 and the 7th non-linear layer Relu7 are equal are as follows:

F (y)=max (0, y)

Wherein, y indicates the input of current layer.

7. Tibetan medicine's urine color automatic recognition method according to claim 1 based on deep learning, which is characterized in that institute State step S3 specifically:

According to the urine color identification model, using softmax function prediction input urine image in nine class urine colors Probability distribution, thus complete Tibetan medicine's urine color based on deep learning automatic identification.

8. Tibetan medicine's urine color automatic recognition method according to claim 7 based on deep learning, which is characterized in that institute State probability distribution p_uExpression formula are as follows:

Wherein, O_nIndicate the linear combination of the n-th output layer, C indicates total urine colour type number, O_uIndicate the line of u output layer Property combination, n indicates that n-th of vector of linear combination, M indicate the first full connection unit FC1, the second full connection unit FC2 and the 1024 dimensional feature vectors that three full connection unit FC3 are generated, x_mIndicate m-th of input mapping of output layer, W_n,mIndicate weight, b_n Indicate the biasing of urine output feature, * indicates convolution operation.

9. Tibetan medicine's urine color automatic recognition method according to claim 7 based on deep learning, which is characterized in that institute It states nine class urine colors and is respectively as follows: yellow, white, red, redness of the skin or complexion, brown, black, green, cyan and blue.

Technical field

The invention belongs to urine identification technology field more particularly to a kind of Tibetan medicine's urine color based on deep learning are automatic Recognition methods.

Background technique

Tibetan medicine is the important component of traditional Chinese medicine, there is more than 2,300 years history, is that Tibetan people passes through length Phase practice, constantly accumulation are improved and what is formed has full theoretical system, the medicine of unique treatment method and strong national characters System.Tibetan medicine's urine examine be by observing the medical technology that diagnoses the illness of Urine in Patients, be the most characteristic diagnostic means of Tibetan medicine it One, belong to prestige, touch, ask observation scope in three big diagnostic methods, has note in the books such as " moon king's medicine examine ", the Four-Volume Medical Code It carries, is had been more than two thousand years of history away from the present.Urine is the product of human body body fluid metabolism, the generation of urine and exclude not only with urinary system It is related, and other vital organs of the human body also assist in wherein.Tibetan medicine thinks that urine is the diet of human body intake by breaking up repeatedly in vivo And generate, urine can reflect the internal informations such as body fever and chills, so the character of urine not only reflects the lesion of urinary system itself, The almost all of pathological change of clinical findings human body all may cause the variation of urine feature, and therefore, Tibetan medicine's urine is examined for body Lesion has important clinical value.

In general, Tibetan medicine's urine process of examining is divided into three phases, i.e., " nine examine when 3 ", successively are as follows: urinates hot stage, observation Color, steam, smell, foam；Thermophase is urinated, suspended things, floating material are observed；It urinates the cool stage, observation urine transformation period, variation Mode and convolution situation.Wherein, the color characteristic for urinating the hot stage is to urinate one of the important feature examined, and urine color is broadly divided into nine Major class is respectively: yellow, white, red, redness of the skin or complexion, brown, black, green, cyan, blue.Urine color characteristic in Tibetan medicine It is the important feature for diagnosing patient's illnesses type, can be used for distinguishing red bar of disease, Baconic's disease, grand disease.When the face of urine Color such as water, color is clear and dilute person is grand type disease；Yellow is red bar of type disease, and white is Baconic's type disease.Total urine test method is from urine Heat, urine temperature, urine cool single stage observe urine feature, and the color for being long placed in rear urine can be varied, and reflect various disease class The color for mainly urinating the hot stage of type.Therefore, the identification of urine color is identified mainly for the color for urinating the hot stage.

Traditional Tibetan medicine's urine color identification is observed by Tibetan medicine doctor and rule of thumb judges dependent diagnostic opinion.Doctor's diagnosis and treatment Experience, level of skill and the environmental condition in the external world etc. can all directly influence last diagnostic as a result, thus subjectivity is strong, can weigh Renaturation is poor, lacks color judgment criteria.Method based on instrument and equipment is easy that, resolution ratio low error loud by environment shadow is big to be lacked It falls into, thus is difficult to accurately identify object color；Machine learning method Classification and Identification color, wherein color histogram is global Color Statistical as a result, therefore having lacked the positional of pixel；The dimension of color moment is less, cannot express seniority top digit well According to；The region of search method that color set uses makes algorithm performance unstable.Conventional machines study is the spy based on hand-designed Sign, this process need artificial judgement color similarity, can ignore a part of feature to a certain extent, cause color identification quasi- True rate is lower；And identify currently based on the color of deep learning both for solid body, also it is not directed to liquid color identification.

Summary of the invention

For above-mentioned deficiency in the prior art, a kind of Tibetan medicine's urine color based on deep learning provided by the invention from Dynamic recognition methods, energy automatic identification urine provide auxiliary for doctor diagnosed and support in the color for urinating the hot stage.

In order to reach the goals above, a kind of the technical solution adopted by the present invention are as follows: Tibetan medicine's urine face based on deep learning Color automatic identifying method, includes the following steps:

S1, acquisition urine image, and each image is labeled；

S2, the urine color identification model according to the urine picture construction based on depth convolutional neural networks CNN；

S3, urine color characteristic is identified according to the urine color identification model, to complete the hiding based on deep learning Cure the automatic identification of urine color.

Further, each image is labeled in the step S1, specifically:

It is yellow, white, red, redness of the skin or complexion, brown, black, green, cyan and indigo plant by urine image labeling collected Chromatic graph picture, and by the size of all urine Image Adjustings to 224 × 224.

Still further, the urine color identification model constructed in the step S2 includes sequentially connected input layer, One convolution unit C1, the second convolution unit C2, third convolution unit C3, Volume Four product unit C4, the 5th convolution unit C5, first Full connection unit FC1, the second full connection unit FC2, the full connection unit FC3 of third and output layer.

Still further, the first convolution unit C1 include sequentially connected first convolutional layer Conv1, it is first non-linear Layer Relu1 and the first pond layer Pool1, the first convolutional layer Conv1 include that size is 11 × 11, and step-length is 96 of 4 Convolution kernel, the step-length of the first pond layer Pool1 are 2；

The second convolution unit C2 include sequentially connected second convolutional layer Conv2, the second non-linear layer Relu2 and Second pond layer Pool2, the second convolutional layer Conv2 include that size is 5 × 5,256 convolution kernels that step-length is 1, and described the The step-length of two pond layer Pool2 is 2；

The third convolution unit C3 includes sequentially connected third convolutional layer Conv3 and third non-linear layer Relu3, institute State third convolutional layer Conv3 include size be 3 × 3, step-length be 1 384 convolution kernels；

The Volume Four product unit C4 includes sequentially connected Volume Four lamination Conv4 and the 4th non-linear layer Relu4, institute State Volume Four lamination Conv4 include size be 3 × 3, step-length be 1 384 convolution kernels；

The 5th convolution unit C5 include sequentially connected 5th convolutional layer Conv5, the 5th non-linear layer Relu5 and Third pond layer Pool3, the 5th convolutional layer Conv5 include that size is 3 × 3,256 convolution kernels that step-length is 1, and described the The step-length of three pond layer Pool3 is 2；

The first full connection unit FC1 includes the sequentially connected first full articulamentum fc1 and the 6th non-linear layer Relu6；

The second full connection unit FC2 includes the sequentially connected second full articulamentum fc2 and the 7th non-linear layer Relu7；

The full connection unit FC3 of third includes the full articulamentum fc3 of third connecting with the output layer.

Still further, the first convolutional layer Conv1, the second convolutional layer Conv2, third convolutional layer Conv3, Volume Four The output of lamination Conv4 and the 5th convolutional layer Conv5 mapAre as follows:

Wherein, M^l-1Indicate the feature quantity of the urine color of input, l and m respectively indicate the rope of the number of plies and urine characteristic pattern Drawing, n indicates the index of filter,Indicate that filter, * indicate convolution operation,Indicate m layers of first of urine input Vector,Indicate the biasing of l layers of n-th of urine output feature.

Still further, the first non-linear layer Relu1, the second non-linear layer Relu2, third linear layer Relu3, Four non-linear layer Relu4, the 5th non-linear layer Relu5, the 6th non-linear layer Relu6 and the 7th non-linear layer Relu7 swash Function f (y) living is equal are as follows:

F (y)=max (0, y)

Wherein, y indicates the input of current layer.

Still further, the step S3 specifically:

According to the urine color identification model, using softmax function prediction input urine image in nine class urine face Probability distribution on color, to complete the automatic identification of Tibetan medicine's urine color based on deep learning.

Still further, the probability distribution p_uExpression formula are as follows:

Wherein, O_nIndicate the linear combination of the n-th output layer, C indicates total urine colour type number, O_uIndicate u output layer Linear combination, n indicate linear combination n-th of vector, M indicate the first full connection unit FC1, the second full connection unit FC2 1024 dimensional feature vectors generated with the full connection unit FC3 of third, x_mIndicate m-th of input mapping of output layer, W_n,mIndicate power Weight, b_nIndicate the biasing of urine output feature, * indicates convolution operation.

Still further, the nine classes urine color be respectively as follows: yellow, white, red, redness of the skin or complexion, brown, black, green, Cyan and blue.

Beneficial effects of the present invention:

(1) present invention is by depth convolutional neural networks automatic learning characteristic from training data, rather than uses by hand The feature of design, the feature learnt identify urine color more effective.In addition, the algorithm proposed is directly in original pixels Upper operation does not need preconditioning technique, saves the time to a certain extent；

(2) deep learning method is applied to by the present invention solves the problems, such as on urine color classification, and energy automatic identification urine exists The color in hot stage is urinated, aid decision is provided for doctor diagnosed and supports

(3) present invention carries out urine picture collection using the urine capture system of autonomous Design, effectively improves urine The efficiency of color identification model training；

(4) it uses step-length for 2 maximum pond in the present invention, largely reduces the space dimension of input urine picture Degree, while the quantity of parameter and weight is greatly reduced, to reduce calculating cost, over-fitting is avoided, improves urine color The generalization ability of identification model；

(5) non-linear layer is utilized in the present invention, while improving the training speed of urine color identification model, is not damaged The accuracy of its model output, and help to alleviate the gradient problem to disappear.

Detailed description of the invention

Fig. 1 is flow chart of the method for the present invention.

Fig. 2 is the structure chart of this depth convolutional neural networks.

Fig. 3 is that Tibetan medicine's urine color identifies block diagram.

Specific embodiment

A specific embodiment of the invention is described below, in order to facilitate understanding by those skilled in the art this hair It is bright, it should be apparent that the present invention is not limited to the ranges of specific embodiment, for those skilled in the art, As long as various change is in the spirit and scope of the present invention that the attached claims limit and determine, these variations are aobvious and easy See, all are using the innovation and creation of present inventive concept in the column of protection.