IEEE VNRVJIET SB in association with IEEE SPS Hyderabad Section, SPS, CIS , CS and Department of ECE are organized a Session on Multiresolution Signal Processing and Machine Learning

#sps #cs #cass #vnrvjiet

IEEE VNRVJIET SB in association with IEEE SPS Hyderabad Section, SPS, CIS , CS and Department of ECE are organized a Session on Multiresolution Signal Processing and Machine Learning on 10th April 2025 from 9 :30 AM to 3:30 PM.

Multiresolution Signal Processing:

This segment provided hands-on experience with multiresolution signal processing techniques, applying theory to practical scenarios.

1. Practical Applications of Multiresolution Signal Processing: Mr. Tarafdar explored the practical use of Discrete Wavelet Transform (DWT) in 1D and 2D, emphasizing perfect reconstruction using available libraries.
2. Hands-on Lab Exercises:
• DWT 1D and Perfect Reconstruction: Participants used PyWavelets (pywt.dwt() and pywt.idwt()) to decompose and reconstruct 1D signals, demonstrating perfect reconstruction with quadrature mirror filter (QMF) pairs.
• DWT 2D and Perfect Reconstruction: The session included applying pywt.dwt2() and pywt.idwt2() to sample brain images, generating four subbands (LL, LH, HL, HH) and verifying reconstruction integrity.
• TFDWT Implementation: Attendees explored the TensorFlow Discrete Wavelet Transform (TFDWT) PyPI package, implementing multilevel DWT and Wavelet Packet Transform (WPT) filter banks for efficient processing.
3. Sample Brain Image Analysis: A sample brain image was used to illustrate 2D DWT applications, providing a tangible example for participants to manipulate and analyze.

Deep Learning Unified with Multiresolution Signal Processing:

This section focused on integrating multiresolution techniques with deep learning, addressing challenges and solutions in a practical context.

1. DWT Layers for CNNs: Mr. Tarafdar demonstrated the creation of differentiable DWT layers using TFDWT, enabling backpropagation in Convolutional Neural Networks (CNNs). Participants tested these layers with sample brain images to enhance feature extraction.
2. Challenges in Integration:
• Shift Variance: Highlighted how small input shifts can destabilize CNN training, a critical issue in dynamic biomedical data.
• Poor Direction Selectivity in 2D DWT: Noted the limitation of ±45° ambiguity in capturing edges, impacting CNN performance.
• Computational Complexity: Addressed the increased cost of multilevel decompositions and the loss of high-frequency details due to downsampling.
3. Solutions:
• Shift-Invariant Transforms: Introduced Undecimated Discrete Wavelet Transform (UDWT) and Dual-Tree Complex Wavelet Transform (DTCWT) to maintain training stability.
• Direction-Aware Representations: Suggested learned masks to improve directional selectivity, enhancing edge detection.
• Optimized Implementations: Utilized TFDWT’s GPU-parallel architecture and wavelet pyramids to retain high-frequency information and reduce computational overhead.
4. Case Studies: Practical exercises reinforced Session 1’s biomedical cases (fetal ECG, brain segmentation, MRI registration), with participants applying solutions to real-world datasets.

Conclusion

Mr. Tarafdar’s hands-on session successfully translated theoretical concepts into practical skills, empowering participants with tools to implement multiresolution signal processing in machine learning. The interactive lab, enriched with TFDWT and ICASSP-2025 research insights, equipped attendees to address real-world challenges, fostering innovation in AI-driven technologies.

• Date: 10 Apr 2025
• Time: 04:00 AM UTC to 10:00 AM UTC
• Add Event to Calendar
  iCal
  Google Calendar

• Vallurupalli Nageswara Rao (VNR) Vignana Jyothi Institute of Engineering and Technology
• Hyderabad, Andhra Pradesh
• India 500090

---

• Contact Event Hosts

• Dr. T. Srinivas | Faculty Coordinator | srinivas_t@vnrvjiet.in

IEEE VNRVJIET SB in association with IEEE SPS Hyderabad Section, SPS, CIS , CS and Department of ECE are organized a Session on Multiresolution Signal Processing and Machine Learning on 10th April 2025 from 9 :30 AM to 3:30 PM.