Convolutional neural network-based classification system design with compressed wireless sensor network images.

Area of Science:

• Computer Science
• Electrical Engineering
• Signal Processing

Background:

• Traditional machine learning algorithms are being replaced by Convolutional Neural Networks (CNNs) for image classification.
• Wireless Image Sensor Networks (WISNs) are crucial for data collection in natural environments but face energy limitations.
• Transmitting high-resolution images from resource-constrained WISNs severely impacts network lifetime due to battery constraints.

Purpose of the Study:

• To propose an energy-efficient pre- and post-processing mechanism for WISNs.
• To reduce data transfer volume while preserving CNN classification accuracy.
• To enhance the operational lifetime of battery-powered embedded devices in WISNs.

Main Methods:

• Implemented an energy-efficient mechanism involving image resizing and color quantization.
• Developed a method for compressing images on embedded devices before transmission.
• Utilized a pre-trained CNN model for classifying compressed image data.

Main Results:

• Achieved a reduction of approximately 71% in transmitted data volume.
• Maintained a high classification accuracy of around 98%.
• Reduced energy consumption by approximately 71% compared to transmitting uncompressed images.

Conclusions:

• Image compression techniques like resizing and color quantization are effective for WISNs.
• Compressed images can be accurately classified by CNNs, enabling energy savings.
• The proposed method significantly extends the network lifetime of WISNs by reducing data transmission demands.