An Enhanced Scheme for Reducing the Complexity of Pointwise Convolutions in CNNs for Image Classification Based on...

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Deep Convolutional Neural Networks (DCNNs) face challenges with parameter bloat in pointwise convolutions.
• Existing methods reduce DCNN complexity but have limitations with divisibility constraints.

Purpose of the Study:

• To propose an improved scheme for reducing pointwise convolution complexity in DCNNs.
• To address divisibility constraints in grouped pointwise convolutions for enhanced efficiency.

Main Methods:

• Utilizing interleaved grouped filters without divisibility constraints.
• Refining the algorithm to handle non-exact divisibility by replicating input channels and allowing varied group filter numbers.
• Implementing a subnetwork with two consecutive convolutional layers (K and L) and an interleaving layer.

Main Results:

• Achieved an 11% reduction in floating-point computations and a 10% reduction in trainable parameters compared to previous methods.
• Demonstrated significant parameter savings on EfficientNet-B0: 76% on CIFAR-10, 89% on Colorectal Cancer Histology, and 91% on Malaria datasets.
• Maintained test classification accuracy across all tested datasets.

Conclusions:

• The proposed scheme effectively reduces DCNN complexity without compromising performance.
• This method offers a practical solution for optimizing DCNNs in image classification tasks, especially on resource-constrained platforms.