PlaneNet: an efficient local feature extraction network.

Area of Science:

• Computer Vision
• Deep Learning
• Mobile Computing

Background:

• Deploying deep neural networks on resource-constrained embedded and mobile devices presents significant challenges.
• Traditional lightweight networks like MobileNetV1 often increase computation time due to extensive use of 1x1 convolutions.

Purpose of the Study:

• Introduce PlaneNet, a novel lightweight convolutional neural network architecture.
• Enhance accuracy while reducing network parameters and multiply-accumulate operations (Madds) for mobile deployment.

Main Methods:

• Developed PlaneNet by optimizing the utilization of 1x1 convolutions for effective local feature extraction.
• Evaluated PlaneNet on image classification (CIFAR-10, Caltech-101, ImageNet2012) and semantic segmentation (VOC2012) tasks.

Main Results:

• PlaneNet achieved higher accuracy (74.48%) than MobileNetV3-Large (73.99%) and GhostNet (72.87%) on classification tasks.
• Demonstrated state-of-the-art performance with fewer parameters and reduced Madds compared to existing models.
• Verified practical applicability for mobile devices due to its efficiency.

Conclusions:

• PlaneNet offers a superior balance of accuracy and efficiency for mobile and embedded deep learning applications.
• The optimized use of 1x1 convolutions in PlaneNet addresses key limitations of current lightweight networks.
• PlaneNet represents a significant advancement towards deploying sophisticated AI models on edge devices.