Learning to Resize Images - PyTorch

PyTorch implementation of Learning to Resize Images for Computer Vision Tasks by Hossein Talebi, and Peyman Milanfar.

This paper proposed an optimized based pre-processing module to resize images.

Usage

class Resizer(in_chs, out_size, n_filters, n_res_blocks, mode)

• in_chs (int) - Number of channels in the input image
• out_size (int or tuple) - The size of images the resizer resize to
• n_filter (int) - Number of output channels in the Resizer's convolution layers. Default: 16
• n_res_blocks (int) - Number of residual blocks in the Resizer. Default: 1
• mode (str) - Algorithm for interpolation. See PyTorch docs for details. Default: 'bilinear'

You can build the resizer on top of your model for different computer vision tasks.

Example:

from resizer import Resizer
class Classifier(nn.Module):
    def __init__(self, ...):
        super(Classifier, self).__init__()
        self.resizer = Resizer(in_chs=3, out_size=128)
        self.model = YourModel
    def forward(self, x):
        x = self.resizer(x)
        x = self.model(x)
        return x

Experiments

Dataset

Food-11

This is a dataset containing 16643 food images grouped in 11 major food categories.

Training

Resize 512x512 to 128x128 using two diffrenet methods and fine tune for 5 epochs.

Result

Model	Resizer	Parameters	Top-1 Acc
vgg11_bn	bilinear	128,816,907	0.87481
vgg11_bn	learnable	128,833,694	0.88646
vgg13_bn	bilinear	129,001,803	0.89423
vgg13_bn	learnable	129,018,590	0.89901
Resnet50	bilinear	23,530,571	0.89512
Resnet50	learnable	23,547,358	0.89961

Reference

Papers

1. Hossein Talebi, and Peyman Milanfar. "Learning to Resize Images for Computer Vision Tasks."

Implementation

1. https://github.com/pytorch