Topic : CNN based Location Image Search and its Adaptation to Social Network, PINPLACE
“CNN based place recognition web app”
- Service of place recognition feature & SNS feature.
- Collect data set & Build CNN models which have the best accuracy
- Work on UI design & graphic Design
- Apply CNN models on web app
The following is a user flow diagram, which shows the connectivity and hierarchy between our web pages.
≪ Cover page / Start Page / User Guide Page / Signup Page / Login Page/
Find Location Page / List Up Page / Upload Picture Page / SNS Page / My Page ≫
-
- Cover page : Since it is the first screen that users face, We designed the logo ourselves because we thought we had to firmly convey the platform brand image.
-
Start Page : This page is expressed in fancy graphics to roughly imply the functionality of our platform
-
User Guide Page : For optimal UX, we made this page with Card UI. Every time user turn the page, the content and design are designed to be different.
-
Sign Up Page : This page is for new users who want to make an account for this service. Currently this includes four text or password boxes, and a submission button.
-
Login Page : To use the service, users need to sign in via this page. Among the information provided in the sign-up page, ID is unique for each user: thus, ID and password are needed to log in. Additionally, there is the button to the sign-up page for who doesn’t have an account for this service.
-
Find Location Page : This page is core function page. We connect with CNN model that we made ourselves.
-
List Up page : This page shows the list of places serviced, by popularity. Popularity can be measured by daily, weekly, or monthly. Each place entry is clickable and shows a subpage for that place.
-
Upload Picture Page : This page is prepared for improving AI model, so the location information for the picture is necessary. The dropdown list for locations needed is served.
-
SNS Page : On the SNS page, you can see recommendations for places shared by celebrities.
-
My Page : This is the own user page for a user logged in, which shows pictures uploaded by that user from find location page.
-
ResNet50 model is adopted
-
Total image data : 25,450
-
Training & validation data: 17,815
-
Input Size : 128 * 128
-
Train set, Validation set, Test set : 5:2:3
-
Classes : 10
-
Batch size : 32 epoch : 80
-
Optimizer : Nadam
Total 10 class
Dongdaemun_Design_Plaza, Gyeongui_Line_Forest_Park, Naksan_Park,
Namsan_Seoul_Tower, The_Hyundai_Seoul_Mall, Myeongdong_Cathedral,
Ikseon_Dong_Hanok_Village, Jamsil_Lotte_Tower, Han_River_Sebitseom,
Haebangchon
- Lenet-5 has three convolution layer, two pooling layer, one fully-connected layer and this have about 60,000 parameter to learn. This model is basic model of CNN.
- AlexNet model has five convolution layer, three pooling layer, two local response normalization layer, one fully-connected layer and this have about 62,000,000 parameter to learn.
- VGG16 model has 13 convolution layer, 5 pooling layer, three fully-connected layer and this have about 138,000,000 parameter to learn. This is much deeper model than AlexNet.
- ResNet model used idea of "skip connection" which solve the gradient vanishing problem which happens when model is deeper. It has 49 convolution layer with pooling layer and one fully-connected layer.
Model | Lenet-5 | AlexNet | VGG16 | ResNet50 |
---|---|---|---|---|
Accuracy | 66.3% | 13.31% | 12.79% | 91.12% |
We Finaly choose ResNet50 model.
- Final Model File (If you want to download the trained model file click this)