A lightweight deep neural network with higher accuracy.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks, particularly lightweight models like MobileNetV2, are crucial for efficient image processing.
• Challenges in deep learning include gradient vanishing/exploding and extracting effective features from complex backgrounds.

Purpose of the Study:

• To design a novel lightweight deep neural network with improved accuracy based on MobileNetV2.
• To enhance feature extraction capabilities and model stability for image classification and object detection tasks.

Main Methods:

• Modified the network depth of MobileNetV2 to balance resolution, width, and depth, stabilizing gradients.
• Introduced a channel attention mechanism into the Bottleneck module to assign channel weights based on feature relevance.
• Implemented a hybrid usage strategy for the improved Bottleneck module within the MobileNetV2 architecture.

Main Results:

• The proposed network demonstrated the highest classification accuracy on ImageNet-1K, CIFAR-10, and CIFAR-100 datasets compared to MobileNetV2, MobileNetV3, ShuffleNetV2, GhostNet, and HBONet.
• YOLOV4-Lite, utilizing the proposed lightweight network, achieved the highest detection accuracy on the PASCAL VOC07+12 dataset.

Conclusions:

• The designed lightweight deep neural network effectively improves accuracy in image classification and object detection.
• The modifications, including depth adjustment and channel attention, contribute to enhanced feature extraction and model stability.