Image Classification in JPEG Compression Domain for Malaria Infection Detection.

Area of Science:

• Computer Vision
• Digital Image Processing
• Medical Imaging

Background:

• Digital images are commonly stored in compressed formats like JPEG.
• Image classification tasks typically require decompression, increasing computational cost and time.
• Classifying images directly within the compression domain offers potential efficiency gains.

Purpose of the Study:

• To investigate the feasibility of image classification directly in the JPEG compression domain.
• To evaluate the effectiveness of compression domain classification for medical images (malaria-infected red blood cells).
• To explore methods for reducing computational overhead in image classification.

Main Methods:

• Analysis of JPEG compression mechanisms.
• Classification experiments using data from various stages of JPEG compression.
• Utilizing Long Short-Term Memory (LSTM) networks for classification.
• Training data included DCT coefficients, DC values (decimal/binary), scan segments, and bitstreams.

Main Results:

• LSTM successfully classified images in their compressed JPEG form with approximately 80% accuracy.
• Classification accuracy exceeded 90% when using only coded DC values.
• Demonstrated that image classes are separable within the JPEG compressed format.
• The compression domain processing method significantly reduced input data size.

Conclusions:

• Image classification is feasible directly within the JPEG compression domain.
• This approach eliminates the need for image decompression, leading to substantial savings in memory and computation.
• Compression domain classification holds promise for efficient medical image analysis and other applications.