The Department of Electronics and Communication Engineering is glad to announce that Dr V Udaya Sankar, Assistant Professor has published the patent (App no. 202141056542), ‘A system and method with Matrix enabled Road distress classification with reduced computational complexity and reduced memory requirements’, in collaboration with Dr Siva Sankar Yellampalli and Ms Gayathri Lakshmi Chinthakrindi.
This work has applications related to visual inspection systems. While this research considers road crack detection application, the same can be extended to various applications such as leaf disease prediction, covid prediction etc. This invention provides an alternative approach instead of using traditional machine learning algorithms that has less computational complexity as opposed to deep neural networks that take more complex operations. This method will also lead to further research in matrix-based machine learning applications related to image processing and image classification.
The research team is planning to collaborate with Efftronics Systems Pvt ltd. for PCB defect detection and discussions are initiated with some start-ups for visual inspection applications. Their future research plan is to look deeper into these algorithms in combination with some of the deep neural networks to reduce computational complexity. In addition, Dr Udaya Sankar is also looking forward to establishing his own start-up in the incubation centre soon.
Abstract of the Research
A method for image classification is provided, wherein, the pre-processed gray scale image is first sent to the feature extraction block, and the said feature extraction block considers every image as a matrix and computes the metrics for features, viz., 1) EMD distance which is popularly known as Wassertain distance/Earth movers distance and is computed with respect to block image and 2) Frobenius Norm which is the square root of the sum of the absolute squares of its elements and finally, 3) Condition Number, which measures the ratio of the maximum relative stretching to the maximum relative shrinking that matrix does to any non-zero vectors. This method is preferred over the existing methods due to the drastic reduction in computational complexities and, utilizing lesser memory. Also, with this method and system, the communicational complexities too are significantly reduced and also, and the results yielded are far more significantly accurate.