Impact of data smoothing on semantic segmentation
- PDF / 2,724,856 Bytes
- 10 Pages / 595.276 x 790.866 pts Page_size
- 59 Downloads / 220 Views
(0123456789().,-volV)(0123456789(). ,- volV)
S.I. : WORLDCIST’20
Impact of data smoothing on semantic segmentation Nuhman Ul Haq1 • Zia ur Rehman1 • Ahmad Khan1 Fawad Qayum3
•
Ahmad Din1 • Sajid Shah1 • Abrar Ullah2
•
Received: 10 May 2020 / Accepted: 2 September 2020 Ó Springer-Verlag London Ltd., part of Springer Nature 2020
Abstract Semantic segmentation is the process to classify each pixel of an image. The current state-of-the-art semantic segmentation techniques use end-to-end trainable deep models. Generally, the training of these models is controlled by some external hyper-parameters rather to use the variation in data. In this paper, we investigate the impact of data smoothing on the training and generalization of deep semantic segmentation models. A mechanism is proposed to select the best level of smoothing to get better generalization of the deep semantic segmentation models. Furthermore, a smoothing layer is included in the deep semantic segmentation models to automatically adjust the level of smoothing. Extensive experiments are performed to validate the effectiveness of the proposed smoothing strategies. Keywords Semantic segmentation SegNet Smoothing Deep learning
1 Introduction
& Ahmad Khan [email protected] Nuhman Ul Haq [email protected] Zia ur Rehman [email protected] Ahmad Din [email protected] Sajid Shah [email protected] Abrar Ullah [email protected] Fawad Qayum [email protected] 1
Department of Computer Science, COMSATS University Islamabad-Abbottabad Campus, University Road, Abbottabad, Pakistan
2
School of Mathematical Computer Sciences, Heriot-Watt University Dubai Campus Dubai International Academic City, Dubai, UAE
3
Department of Computer Science, IT University of Malakand, Chakdara, KPK, Pakistan
Dividing an image into non-overlapping homogeneous regions is called segmentation [35–39] but semantic segmentation is the process to divide an image into regions such that each region contain a holistic object [15, 25, 25]. Besides other applications [3, 26, 57, 58] the deep neural networks have been successfully used in semantic segmentation [24, 42, 74] to classify images at pixels level. The deep end-to-end trainable networks have replaced the manual features extraction, which was a primitive step in classical machine learning paradigm. It is important to increase the generality of the these networks and reduce the training overhead. Different architectures [10, 14, 25, 31, 40, 51, 59, 62], hyperparameters selection and tuning [5, 16, 60, 65, 76], and data augmentation [13, 20, 56, 63] are considered to achieve the mentioned goals. The human visual processing system is divided into a sequence of hierarchical stages to analyze the visual input. It uses a particular region in brain to perform a particular task. Field et al. [23] studied the selective nature of brain neuron in term of orientation and spatial frequency. Similarly, Watson et al. [71] investigated the human brain scene-selective nature by presenting indoor and outdoor scene imag
Data Loading...
