DO-Conv

DO-Conv augments convolutional neural networks by introducing a depthwise over-parameterized convolutional layer that increases representational capacity during training to improve performance on vision tasks such as image classification, detection, segmentation, and bioinformatics imaging tasks including protein structure prediction and cellular imaging.

Key Features:

• Depthwise Over-Parameterization: Incorporates an additional depthwise convolution alongside the standard convolution, convolving each input channel with a unique 2D kernel to increase the number of learnable parameters without altering the linear nature of the operation.
• Training Replacement: Replaces conventional convolutional layers during the training phase to provide augmented capacity for learning complex patterns.
• Inference Integration: Integrates the depthwise convolution into the conventional convolution layer at inference so the computational load matches that of a standard convolutional layer.
• Framework Implementations: Implementations are provided for TensorFlow, PyTorch, and GluonCV.
• Performance Enhancement: Demonstrated through experimentation to improve CNN performance on classical vision tasks including image classification, detection, and segmentation.

Scientific Applications:

• Computer Vision: Improves model accuracy in image classification, object detection, and image segmentation benchmarks.
• Bioinformatics Imaging: Enhances CNN-based analyses for tasks such as protein structure prediction and cellular imaging where higher representational capacity during training can improve results.

Methodology:

During training, DO-Conv replaces standard convolutional layers with a composite layer that adds a depthwise convolution applying a unique 2D kernel per input channel; during inference, the depthwise component is merged into the conventional convolution so computation is equivalent to a standard convolutional layer.