Edge Detection

Project Title

Edge Detection

Project Area of Specialization Computer ScienceProject Summary

Edge Detection is a key low-level task in computer vision which can greatly help in several high-level computer vision tasks including surveillance, panoptic segmentation, object detection, etc. Accurate edge detection is critical in several applications including self-driving cars and medical images. There are various methods available for edge detection ranging from traditional hand filters to deep learning edge detection. However, these methods are limited in their scope and suffer from dataset bias. In this project, we will study the available method on edge detection and then design novel methods for edge detection based on deep learning. We will also collect and synthesize new dataset which will provide us with sufficiently varying dataset to test the strength of our algorithm. Furthermore, we will test our methods on medical images and determine their performance in the detection of tumors particularly for cancer detection. The techniques could also be applied for COVID detection by analyzing the images of lungs.

Project Objectives

Our objectives for this project are as follows:

• We will explore existing edge detection methods.

• We will build a new dataset with less bias.

• We will build a new method for edge detection.

• We will evaluate existing methods on our dataset.

• We will evaluate our method on existing datasets.

• We will apply our methods on medical imaging problems, particularly for tumor detection in cancer patients

Project Implementation Method

In this project, we will propose a deep learning method for edge detection and a synthetic textured dataset.We will train our proposed method on synthetic textured dataset and BSDS500 (The Berkeley Segmentation Dataset) separately till validation loss converges. We will compare our results with other state-of-the-art edge detection methods such as BDCN, DexiNed, and HED on both datasets, and additionally with RCF on Synthetic dataset. All these methods will also be trained on relative datasets until their validation loss converges.

Benefits of the Project

Edge detection is one of the most fundamental computer vision tasks which can be used to aid a variety of mid-level and high-level computer vision tasks such as image segmentation, object recognition and object detection. Even though a significant amount of work has been done in tackling this problem, it is still an open problem due to biases in the datasets and methods proposed. Therefore, we want to further explore this problem and propose a new dataset and a method while minimizing the bias. More critically, we would use the develop methods on medical images for cancer detection or size of tumors.

Technical Details of Final Deliverable

The previous methods used for edge detection have limitations. There is a lack of consistency when it comes to detecting edges in a variety of datasets. They are trained to find edges on datasets with specific underlying distribution, if we change underlying distribution, performance of these techniques drops significantly. Furthermore, convolution neural networks (CNNs) proposed in above methods are highly complex. Most of these CNNs have parameters in the order of 10 Million.

In this project, we will propose a Convolutional Neural Network, which will work equally well on edge detection datasets with different underlying distributions. We believe edge detection is a non-complex task therefore our proposed CNN will be much less complex compared to other methods but we intend to obtain similar or improved results compared to previous methods. Furthermore, we will also introduce diverse and consistent synthetic datasets, which will be more challenging than traditional datasets.

Final Deliverable of the Project Software SystemCore Industry ITOther Industries Medical , Transportation , Media , Health , Telecommunication Core Technology Artificial Intelligence(AI)Other Technologies Augmented & Virtual Reality, RoboticsSustainable Development Goals Industry, Innovation and InfrastructureRequired Resources

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	26000
Google Colab pro Account	Equipment	3	2000	6000
External hard disk	Equipment	1	10000	10000
travel for data collection	Miscellaneous	1	10000	10000