Start to reply:
The direction of our project is image recognition based on deep learning. Image recognition can be said to be a very basic and promising technology in artificial intelligence.
Computer image recognition technology has applications in many fields such as public safety, biology, industry, agriculture, transportation, and medical treatment.
For example, the license plate recognition system in transportation; face recognition technology and fingerprint recognition technology in public safety; seed recognition technology in agriculture, food quality detection technology; electrocardiogram recognition technology in medical field, etc. With the continuous development of computer technology, image recognition technology is also continuously optimized, and its algorithm is also continuously improved.
Our team implemented an image recognition system using the ResNet model in machine learning and the existing cifar10 database. After our testing, the accuracy of the system for images was as high as 90%. Due to the lack of equipment, our network improvement is very slow It is also very laborious, and it is considered unsatisfactory to achieve this result.
Below I will respond to the two aspects of network model and data set selection.
The network model we used was ResNet proposed by three Chinese scholars in 2015. This model has the following advantages over other networks.
A deep residual learning framework is introduced. Instead of letting some stacked layers learn the original features directly, the stacked layers are used to fit the residual map.
ResNet can get an increase in accuracy from the increased depth, and the results produced by Resnet have also refreshed the record at the time.
ResNet has strong scalability, and when the network is increased to 100 or even 1000 layers, there will still be no degradation problems.
So here we use this network to realize image recognition technology.
Let me introduce the cifar10 data set, which is composed of 60,000 32x32 color images in 10 categories, with 6000 images in each category. There are 50,000 training images and 10,000 test images.
The training set is a small data set collated by Alex and Ilya for identifying universal objects. The training set is universal and practical. So we take this data set.
After the experiment, we also got better training results.
The system is also very simple to use. Store the picture in the img folder and name it as name.jpg, and then run the predict.py program directly to get the result.
This system is the first project of our team. After seeing the result, we are ecstatic and hope to get everyone's approval.
The following video demonstrates the actual measurement effect of the system.