2:00 PM - 2:20 PM
[4P3-J-10-01] Application of Memory Reduction into Environmentally Invariant Regressor CNN
Keywords:image recognition, deep learning, low-rank approximation
In recent years, the need for computer vision applications such as object/scene classification has grown rapidly.
For example, tracking of vehicles and recognizing number plates can be used to automatically estimate traffic flow.
These applications require high environmental invariance (due to weather conditions, illumination changes, etc.) as well as compact design and fast rocessing speed in order to be deployed into edge devices.
In this paper, we consider the regression problem for vehicle's number plate recognition.
Recognition accuracy increases if letters/digits are geometrically transformed beforehand to fix distortions or mis-alignments.
However, using CNN to realize such regression task requires a large nunber of parameters.
We propose to use low-rank approximation and fine-tuning to reduce the size of such CNNs.
Experiments show that under various environments, our method is effective in greatly reducing memory consumption without significantly degrading performance.
For example, tracking of vehicles and recognizing number plates can be used to automatically estimate traffic flow.
These applications require high environmental invariance (due to weather conditions, illumination changes, etc.) as well as compact design and fast rocessing speed in order to be deployed into edge devices.
In this paper, we consider the regression problem for vehicle's number plate recognition.
Recognition accuracy increases if letters/digits are geometrically transformed beforehand to fix distortions or mis-alignments.
However, using CNN to realize such regression task requires a large nunber of parameters.
We propose to use low-rank approximation and fine-tuning to reduce the size of such CNNs.
Experiments show that under various environments, our method is effective in greatly reducing memory consumption without significantly degrading performance.