Summary of the invention
It is an object of the invention to overcome the control of kitchen ventilator in the prior art not ideal enough, the bad technology of smoking effectProblem provides a kind of range hood control method based on image recognition, control system, range hood.
To achieve the above object, The technical solution adopted by the invention is as follows: a kind of range hood control based on image recognitionMethod processed, which comprises
The cooking environments image generated in cooking process is obtained in real time；
The oil smoke situation under cooking environments is judged according to the cooking environments image；
The control instruction of range hood is determined according to the oil smoke situation；
The working condition of range hood is controlled according to the control instruction.
Further, convolutional neural networks model is established, by trained convolutional neural networks model to the culinary artAmbient image carries out identifying processing.
Further, by extracting the characteristic image of cooking environments image to cooking environments image progress process of convolution,Classify after carrying out batch regularization, mapping, pondization processing to characteristic image again to pixel each in characteristic image, is divided into oilCigarette pixel and non-oil smoke pixel.
Further, oil smoke situation is obtained after carrying out statistics calculating to oil smoke pixel and non-oil smoke pixel.
Further, the control instruction includes power-on instruction, when oil smoke situation reaches preset threshold, controls fume-exhaustingMachine is opened.
Further, the control instruction includes air quantity regulating command, and the air quantity of range hood is adjusted according to oil smoke situation.
A kind of range hood control system based on image recognition, including
Module is obtained, for obtaining the cooking environments image generated in cooking process；
Judgment module, for judging oil smoke situation according to the cooking environments image；
Analysis module, for determining the control instruction of range hood according to the oil smoke situation；
Control module, for controlling the working condition of range hood according to the control instruction.
It further, further include study module, the study module is used for after cooking environments image study trainingCarry out identifying processing.
Further, the study module includes convolutional neural networks model.
A kind of range hood further includes Image Acquisition including the above-mentioned range hood control system based on image recognitionDevice, described image acquisition device is connect with the range hood control system signal based on image recognition, for acquiring fume-exhaustingCooking environments image near machine is simultaneously sent to acquisition module.
Further, described image acquisition device includes the camera being set on kitchen ventilator.
By the above-mentioned description of this invention it is found that compared with prior art, one kind provided by the invention is based on image recognitionRange hood control method, control system, range hood, obtain the cooking environments image that generates in cooking process in real time；RootThe oil smoke situation under cooking environments is judged according to the cooking environments image；The control of range hood is determined according to the oil smoke situationInstruction；The working condition of range hood is controlled according to the control instruction, realizes range hood with the automated intelligent of oil smoke situationControl, range hood smoking effect is more preferable, not only makes the more environmentally-friendly energy conservation of range hood, also reduces user in cooking processThe redundant actions for adjusting range hood gear, promote the usage experience of user.
Specific embodiment
The technical solution in the present invention is clearly and completely retouched below with reference to the attached drawing in the embodiment of the present inventionIt states, it is clear that the described embodiments are merely a part of the embodiments of the present invention, instead of all the embodiments.Based on the present inventionIn embodiment, every other implementation obtained by those of ordinary skill in the art without making creative effortsExample, should fall within the scope of the present invention.
It should be noted that term " includes " of the invention and " having " and their any deformation, it is intended that coverCover it is non-exclusive include, for example, the process, method, system, product or equipment for containing a series of steps or units are not necessarily limited toStep or unit those of is clearly listed, but may include be not clearly listed or for these process, methods, productOr other step or units that equipment is intrinsic.
As shown in Figure 1, a kind of range hood control method based on image recognition, the described method comprises the following steps:
S1: the cooking environments image generated in cooking process is obtained in real time；
S2: the oil smoke situation under cooking environments is judged according to the cooking environments image；
S3: the control instruction of range hood is determined according to the oil smoke situation；
S4: the working condition of range hood is controlled according to the control instruction.
In the present embodiment, it using image collecting device such as camera captured in real-time or timed shooting, obtains user and is cookingThe cooking environments image generated during preparing food, such as the cooking environments image in acquisition kitchen, compartment, can be according to acquisition environmentThe camera of different number is arranged in size, such as a camera is arranged in common household kitchen,
After the cooking environments image got, cooking environments image is identified, the oil smoke in image is identified, in oil smokeIdentification during this, the present embodiment is using convolutional neural networks model, as shown in Fig. 2, convolutional neural networks modelBy input layer (Input images), convolutional encoding network (code module), deconvolution decoding network (decodeModule), output layer (output images) forms.Wherein, Input images is used to the image of acquisition being input to modelIn, code module is the full convolutional neural networks of a multilayer, is all connected to down-sampling layer after every layer of convolutional layer, the convolutional encodingNetwork structure is similar to the preceding 16 layers of convolutional Neural layer of VGG-19 network structure designed for object classification, while abandoningThe full articulamentum of VGG-19 is conducive to export high-resolution characteristic pattern in the encoder of bottommost layer in this way, and reduces netThe parameter of network, to reduce the training time of network.Corresponding deconvolution decoding network decode module also has 16Convolutional layer, so whole network structure has reached 32 layers, output images is made of a classifier (Softmax), is usedThe pixel of corresponding position is classified as each classification, calculates the probability for belonging to which classification, the identification for oil smokeFor, input image pixels are mainly divided into two classes: oil smoke pixel, non-oil smoke pixel,
For cataloged procedure as shown in figure 3, in convolutional encoding network, convolution operation each time is all by big with 3x3The convolutional layer (Conv) of small convolution kernel carries out feature extraction to the output image of up-sampling layer (Upsamleing), then to instituteThe feature of extraction carries out batch regularization operation (Batch Normalization, abbreviation BN) in batch regularization layer, and utilizes activationFunction (ReLu) carries out Nonlinear Mapping to feature, and the Dropout layers of node made in network is recycled to close with certain probabilityIt closes, finally carries out pondization operation (Pooling) in pond layer, maximum pond (Max-Pooling), Chi Huahou are used in this implementationThe length and width of each feature become original half, and the maximum available image of pondization is flat in the variation of small space displacementMotion immovability, and multiple maximum pondizations can obtain the feature of more robust property for classifier (Softmax), it is continuous to carry outPond down-sampling will cause picture and constantly be distorted, and boundary information is lost, and be unfavorable for the segmentation task of image, so correspondingIt is subsequent setting decoding network, in order to restore image information as far as possible, it is also necessary to be recorded during pond maximum specialThe index position of value indicative.Decoded process is as shown in figure 4, each decoder uses the aspect indexing of pond process record to inputFeature is up-sampled, reuse with can training convolutional core convolutional layer (Conv) to up-sampling layer (Upsamleing) outputSparse features figure carry out convolution operation and obtain dense characteristic pattern, the process is similar with cataloged procedure, and then carries out batch justThen change operation (Batch Normalization, abbreviation BN), activation primitive ReLu Nonlinear Mapping, pondization operation(Pooling), output layer (output images) is that each pixel is classified using classifier (SoftMax), outputIt is the corresponding class probability of each pixel, the classification of the maximum probability of each pixel is exactly the classification predicted, in the network modelIn structure, in order to overcome deep neural network to be difficult to trained disadvantage and accelerate the training process of network, after each convolutional layerAddition batch regularization layer, it is therefore prevented that the gradient disappearance problem that depth network is easy to appear in the training process helps to improve instructionExperienced convergence rate and model accuracy, while network is designed using Dropout layers to prevent model from the phenomenon that over-fitting occurAfter structure, need to be trained it to obtain prediction model.Prepare data set for training process, acquires some image making numbersAccording to collection, data set claps the image for taking camera to install under visual angle, includes oil smoke and not no oil smoke, and carry out hand to data setWork mark, is divided into 2 kinds of semantic classes: oil smoke, non-oil smoke for each pixel, reads data in order to facilitate computer, marks respectivelyIt is denoted as 0 (non-oil smoke) and 1 (oil smoke).Can be obtained by the parted pattern of oil smoke by training sample, can by input picture intoRow segmentation, its pixel is classified, oil smoke pixel, non-oil smoke pixel are divided into,
All pixels are carried out oil smoke situation is calculated, such as the number of pixels of oil smoke is m, non-oil smokeNumber of pixels is n, then oil smoke size P=m/ (m+n), and the value range of P is [0,1],
As shown in figure 5, for range hood set a booting threshold value k, as P<k, range hood remains turned off state, work as P>When k, i.e., when oil smoke reaches certain oil smoke concentration, power-on instruction is generated, control range hood is opened, and at this moment range hood startsWork can also be adjusted, according to the size of P value after range hood work according to wind-force of the oil smoke concentration to range hoodAnd air quantity regulating command, such as the revolving speed by controlling motor are generated according to oil smoke concentration, reach the control to range hood wind-forceSystem, and then the air quantity of range hood is adjusted, the both hands of user can be liberated during user cooks in this way, improve userUsage experience.
The present embodiment additionally provides a kind of range hood control system based on image recognition, as shown in fig. 6, including obtainingModule 1, judgment module 2, analysis module 3, control module 4, study module 5,
The module 1 that obtains obtains the cooking environments image generated in cooking process, and the judgment module 2 is cooked according to describedAmbient image of preparing food judges oil smoke situation；The analysis module 3 determines the control instruction of range hood according to the oil smoke situation；InstituteState the working condition that control module 4 controls range hood according to the control instruction；The study module 5 is to the cooking environmentsIdentifying processing is carried out after image study training, which is convolutional neural networks model, specifically may refer to above-mentioned be based onCooking environments image is divided into oil smoke by convolutional neural networks model by the description in the range hood control method of image recognitionIt after non-oil smoke, is calculated, judges oil smoke state by calculating, the working condition of range hood is controlled according to oil smoke state,For example, be that range hood sets a booting threshold value according to oil smoke concentration, when oil smoke concentration is less than booting threshold value, range hoodState is remained turned off, when oil smoke concentration is greater than booting threshold value, control range hood is opened, and at this moment range hood is started to work,It after range hood work, can also be adjusted according to wind-force of the oil smoke concentration to range hood, such as pass through control motorRevolving speed reaches the control to range hood wind-force, and then adjusts the air quantity of range hood.
The present embodiment additionally provides a kind of range hood, as shown in fig. 7, comprises the above-mentioned fume-exhausting based on image recognitionMachine control system 100 further includes image collecting device 200, described image acquisition device 200 and the fume-exhausting based on image recognitionThe connection of 100 signal of machine control system obtains module 1, institute for acquiring the cooking environments image near range hood and being sent toStating image collecting device includes the camera being set on kitchen ventilator, which can be set on kitchen ventilator, for example, settingIt at the suction opening of range hood, alternatively, the marginal position of range hood is arranged in, or can be set on hearth, with suctionThe connection of kitchen ventilator signal, camera use the above-mentioned range hood control method based on image recognition to oil suction after taking imageSmoke machine is controlled.
By the above-mentioned description of this invention it is found that compared with prior art, one kind provided by the invention is based on image recognitionRange hood control method, control system, range hood, obtain the cooking environments image that generates in cooking process in real time, adoptIt is identified after carrying out learning training with convolutional neural networks model, the oil smoke under cooking environments is judged according to the cooking environments imageSituation；The control instruction of range hood is determined according to the oil smoke situation；The work of range hood is controlled according to the control instructionMake state, range hood can be enabled according to the automatic start and stop of oil smoke situation, according to the adjustment wind-force of oil smoke concentration intelligenceSize, realize range hood with oil smoke situation automated intelligent control, range hood smoking effect is more preferable, not only makes oil suctionThe more environmentally-friendly energy conservation of smoke machine also reduces user in the redundant actions of the adjustment range hood gear of cooking process, promotes user'sUsage experience.
It above are only several specific embodiments of the invention, but the design concept of the present invention is not limited to this, all benefitsIt is made a non-material change to the present invention, should all be belonged to behavior that violates the scope of protection of the present invention with this design.