DCTCNet: Sequency discrete cosine transform convolution network for visual recognition.

Area of Science:

• Computer Vision
• Deep Learning
• Signal Processing

Background:

• Discrete Cosine Transform (DCT) is vital for image compression and quality but suffers from blocking effects.
• Blocking effects in conventional DCT limit its efficiency in computer vision applications.
• Addressing these limitations is crucial for advancing vision tasks.

Purpose of the Study:

• To introduce a novel DCT variant, All Phase Sequency DCT (APSeDCT), into convolutional networks.
• To develop an APSeDCT-based convolutional module (APSeDCTConv) for extracting multi-frequency information.
• To propose an augmented convolutional operator (MultiConv) using APSeDCTConv to improve model efficiency and performance.

Main Methods:

• APSeDCT was integrated into convolutional networks, creating APSeDCTConv, which mimics vanilla convolution.
• A MultiConv operator was designed by augmenting APSeDCTConv.
• The last three bottleneck blocks of ResNet were replaced with MultiConv for experimentation.

Main Results:

• APSeDCTConv demonstrated transferability similar to standard convolutions.
• Replacing ResNet blocks with MultiConv reduced computational costs and parameter count.
• Consistent performance gains were observed in image classification on ImageNet across various models (ResNet, Res2Net, ResNext).
• Object detection and instance segmentation on COCO showed 0.5%-1.1% and 0.4%-0.7% AP improvements, respectively.

Conclusions:

• APSeDCTConv effectively extracts multi-frequency information from deep features.
• The proposed MultiConv operator enhances model efficiency and performance in computer vision tasks.
• APSeDCTConv augmentation offers a promising approach for improving deep learning models in vision applications.