This research investigated captioning papers based on the attention mechanism and replicate the result of the latest research. The performance of each
algorithm was tested on the data set.
Explored the mechanism of RNN , implement seq2seq model.
Replicated the results of the paper which use the memory module to memorize the knowledge about language style, and implement the generation of
image captioning text in a specific language style.
Investicated attention based image captioning mechanism, using encoder-decoder in the form of weighted local visual features to generate a
description. The survey results show the attention widely thought to use, can not only add it to decoder, but also encoder.
The current image description opens up many new research directions, such as adding style elements. Current datasets are mainly trained and tested
based on natural images, and it is expected that image description methods for remote sensing images and even medical images will be proposed in the
Tools : MATLAB, Python
Classical statistical machine learning and graph probability models are used to process image data, and various supervised and unsupervised algorithms
are implemented. The tests are performed on real data sets, and the algorithm performance is compared and analyzed.
Image data pre-processing. Programs in MATLAB are used to implement image coordinate conversion, gray mapping, histogram correction, and spatial filtering, and test on real data sets.
Target Detection. SIFT corner detection technology is implemented in Python and tested on real images.
Image pattern recognition. The Back Propagate Network and CNN was implemented to classify the handwritten data set MNIST . After PCA dimensionality reduction, optimization of the loss function and learning rate, the classification accuracy rate of 97.3% by BP Network adn 99.2 by CNN was finally
Compared the classification differences of SVM model, Tree Model and CNN on cat and dog dataset(on Kaggle). It is found that when the sample data volume reaches 3000 , the classification accuracy of SVM and Tree Model decreases significantly, and the classification effect of CNN network is stable.
This project originated from an observation with my partners: In the university's teaching buildings,
libraries, and other public places, it is often the case that the room is empty but the air conditioner is still
running. We hope to propose a thermostatic system to avoid such a waste of resources.
This project proposes an intelligent indoor constant temperature control system, which can automatically
control heating, ventilation and air conditioning equipment (such as air conditioners, electric heaters, etc.), and keep the indoor temperature at a comfortable temperature under the condition of the lowest energy
This project originated from an observation with my partners: In the university's teaching buildings, libraries, and other public places, it is often the case that the room is empty but the air conditioner is still running. We hope to propose a thermostatic system to avoid such a waste of resources.
This project proposes an intelligent indoor constant temperature control system, which can automatically control heating, ventilation and air conditioning equipment
(such as air conditioners, electric heaters, etc.), and keep the indoor temperature at a comfortable temperature under the condition of the lowest energy consumption. We propose the following design schemes:
1. Intelligent temperature control system built many types of sensor applications , it can continuously monitor the indoor temperature, humidity changes in the
surrounding environment, lighting and thermostat, such as it can be determined whether any of the room (if there is movement) To determine whether to turn on the temperature adjustment device.
2. Intelligent temperature control system has the ability to learn, as you every once in a set time setting up a certain temperature, it will be recorded once, and then after a long period of training , it will be able to learn and remember the user's daily routine and habits Temperature preferences, and it will automatically generate a setting
solution using algorithms. As long as your lifestyle does not change, you no longer need to manually set the intelligent constant temperature controller .
3. Intelligent temperature control system also supports networking functionality , which means you can use its remote control handset, but its use is very simple, just
install it on the wall to adjust the way it is cool, only Just turn its control knob .
4. The air-conditioned indoor cascade of different locations to achieve a constant temperature control system air-conditioning system for the entire family. In order to
optimize the regulation of energy consumption, while maximizing user comfort.
1. The Optimum Temperature
Having our thermostat automatically combine temperature, humidity, pollen, air pollution levels, and personal preferences for temperature, humidity, and air purity is a
challenge. Because even in a room where only one person lives, the optimal temperature varies with multiple variables.
In the above background, we use the PMV-PPD indicator to predict the optimal temperature of the indoor human body.
The PMV indicator is based on the thermal comfort equation and combines all variables to more comprehensively evaluate the environmentally induced thermal
sensation and has been incorporated into the IS07730 standard.
The following is the calculation method of PMV-PPD index.
The discrepancy between heat dissipation and external heat dissipation
The human body metabolism of unit skin area
The external activity
The clothing thermal resistance
The ratio of the surface area of the human body covered by the clothing to the surface area of the human body exposed to the outside
The air temperature
The average radiant temperature
The relative air speed
The clothing surface temperature
The convective heat transfer coefficient
The water vapor pressure around the human body
For different indoor and outdoor environments, the optimum temperature may vary. We use the RBF neural network to predict the optimal indoor temperature based on a set of PMV index data collected over a period of time .
Figure 1 Structure of RBF neural network
In constructing the RBF network model, X = [x1，x2，...，x6]T is the input vector of the network, that is, the thermal resistance of clothing , body metabolism per unit
skin area , air temperature , average radiation temperature , relative Air velocity , water vapor pressure around the human body ; Y is the output variable, i.e. the PMV
index ; H = [h1，h2，...，hn]T is the radial basis vector of the network.
A total of 1,000 data samples were used to train the network . After training , in order to test the computational accuracy of the network, we selected 14 sets of untrained data samples to test the network.
The test results based on the RBF neural network are shown in the figure. The abscissa indicates the number of samples, and the ordinate indicates the actual and
predicted values of the PMV indicator.
Figure 2 Predictive effect on PMV indicators
Simulation results show that the PMV index based on RBF neural network has high accuracy, and the predicted value can well estimate the true value of PMV index.
This greatly reduces the risk of thermostat signal delays.
When we have a series of historical PMV indicators, and use algorithms to generate future changes in PMV indicators, we will use this indicator to control indoor
environmental conditions and make conditions such as temperature tend to the most comfortable values.
The thermostat affects the air temperature and humidity of the indoor heating / cooling equipment. These are the two main factors affecting the PMV indication. When
we predict PMV> 0 , it means that the human body feels warmer. At this time, our thermostat will affect the equipment cooling, increase the indoor air circulation speed
or increase the indoor relative humidity ; on the contrary, when we predict PMV <0 , this means that the human body feels cold. At this time, our thermostat will reduce
indoor air circulation speed or appropriately reduce indoor relative humidity to adjust indoor comfort. In this way, the optimal temperature adjustment function of the
room is completed.
2. Home Self-starting Preheating and Shutdown Functions
Our home devices need to know the user's activities in advance, so that the warm-up time can be reserved for heating / cooling in the most energy-saving way .
We need to perform long-term observation and measurement while deploying a large number of sensors. The practical work has not yet been implemented. The current theoretical design is as follows:
We expand our design description with the following example measurement data.
We found daily activity data for a female user throughout 2010 who stayed at home alone. The data is recorded by sensors deployed in a woman's homes.
Figure 3 Sensor deployment in a lady's home
User activity is indicated by sensor tags in the data set, including eating, washing, sleeping, going out, getting in, etc. Obviously, here we only need the data labeled
"Leave Home" , "Home " , "Sleep Event", so we use batch script to extract the data. In the RBF neural network model, we use the departure time as an independent
variable and the belonging time as the dependent variable. We select 431 sets of historical data of sensors as sample data for network construction training, and use
the remaining 31 sets of data for network inspection. We obtained a graph of the relationship between such predicted data and actual data.
Figure 4 Predicted Arrived Time
Through the prediction graph, we can see that the RBF neural network model can basically predict the moment when the user leaves home.
However, considering that the user may leave the home for a short time, it is a problem to turn off the thermostat control immediately when the user leaves the home.
Because if the user goes out for a short time, the frequency of closing and opening the thermostat control will increase over time, which is more energy consuming for
To avoid this, we have to calculate and predict when the user will leave home. In this way, we can set a threshold and hold it for 5 minutes. If the time to leave home
(time to return home minus time to leave home) is greater than the threshold, the thermostat control is turned off, and the thermostat control is maintained when the
threshold is less than or equal to the threshold.
We also established an RBF neural network model, using the same sample data, selecting 431 sets of historical data from the sensors as sample data for network
construction training, using the last 31 sets of data for network testing, and obtaining the predicted time to leave home and leave home Diagram of running away. The
chart is as follows.
Figure 5 Prediction of user's departure time
By predicting the time of departure and arrival, our home system can warm up or shut down the temperature system some time in advance to achieve the best energy
saving effect. In this way, the function of intelligent startup and shutdown of the constant temperature system is achieved .
" In addition to posing easily, you can also take photos with realistic virtual characters across time and space . "
In the era of the mobile Internet, images have become an important medium for disseminating information, and the emergence of generations of camera software has
greatly satisfied people's needs for recording and sharing lives. Based on market research, our team has co-designed a new generation of camera software . We
mainly solve the following problems:
- Unique and beauty are most concerned about the photographer
- different places makes it impossible to group photo with family and friends
-There is no easiest way to edit the photo you want
Our solution is:
Design three camera modes: portrait shooting mode, model shooting mode and photo editing mode. They can recommend personalized poses, let users take photos
with realistic virtual characters, and powerful editing and photo scoring functions.
Imagine that when you take a photo with a partner or friend, you will always hear the cameraman's chattering instructions: lower your head a bit or get closer, but the
result of the shot is often not consistent with the effect you expected. Often frustrating. Our portrait shooting mode will intelligently provide users with satisfactory
shooting options at this point: when a character walks into the viewfinder, he can choose a preset shooting style, such as: romantic, witty, solemn, etc., the portrait
shooting mode will automatically Analyze the surrounding scenes and provide a pose recommendation model suitable for the user from the database. In this way, users can more easily adjust their posture and take satisfactory photos.
Figure 1 Portrait mode
We design interesting model shooting modes to " close " each other. When couples cannot meet for a long time due to work or study, it can provide authorized virtual
couples to take pictures with you in this special way to create the joy of couples meeting and record important moments. The operation is very simple. The first step is
to enter a realistic virtual three-dimensional character model that matches the real person in the software. The second step is to take the camera's viewfinder frame to
the position you want to shoot, and the previously entered model will appear In the photo frame, this is AR technology; finally you can walk into the lens and take a
photo with the model (couple)!
Figure 2 Shooting with Model mode
This software will make it easier for users to pose during shooting, and gain happiness and fun in photos of AR virtual characters.
As the person in charge of this project, I led this project to achieve second place in the business plan competition of the class.
“ Sketchpad ” Drawing App on Win32 console
In VS2015 under development environment, I use computer graphics and Windows programming knowledge, to achieve a user-friendly, full-featured, easy to use the
brush software. The brush software can provide the following functions:
1. Basic graphics drawing: rectangle, polygon, ellipse, circle, etc.
2. Interactive curve drawing: Provide Bezier curve drawing function, can call graphic function, and have multiple control point parameter selection.
3. Graphic fill: two graphic fill algorithms can be provided - scan line fill algorithm, flood fill algorithm to fill primitives, and allow users to set.
4. Graphic file management functions, including new, open and save functions.
5. Eraser function to erase graphics on the drawing board.
Brush main interface as Figure 1:
Figure 1 Main Interface
The top of the interface settings Menu , has the following sub-menu " File " , " display " , " graphic " , " fill " , " tool " , " Help " , respectively, the following functions:
" File " provides new, save, save as, open, and exit options;
“ Display ” provides line style and color options. Among them, the line style option can select different pen thicknesses and dots. The color option can open the color plate to freely choose the desired color.
" Graphics " provides five types of drawing primitives, which are rectangle, polygon, circle, ellipse and cubic Bezier curve. Click this menu bar button to provide a
visual single choice "·" logo;
" Filling " provides two filling methods that users can choose - scan line filling and flood filling;
" Tools " provide tools for eraser and empty artboard;
" Help " to provide software-related copyright information.
(See Figure2 )
Figure 2 Menu Entries
I used to write greedy snakes in C ++ . At that time, the window was completely used by the graphics library function to draw messages. The response was completely
completed by timers and button monitoring. It felt like a game completed frame by frame. And this project uses the MFC framework, which greatly reduces my program
ming workload. Many useful message functions are all encapsulated. This can put a lot of energy on the design of algorithms and software functions.
MFC is a convenient application development framework that requires only a low level of knowledge to complete the development of simple applications.
In thinking about the process of drawing and filling primitives, I feel that the graphics algorithms are vast and profound. With this software, you can use graphics to
design graphics and be able to interact with each other. I realized that graphics is now playing a huge role in computer-aided design. This project has increased my