Journal of the Institute of Industrial Applications Engineers Vol.3, No.1, pp.15–23, (2015.1.25) DOI: 10.12792/JIIAE.3.15 Online edition: ISSN 2187-8811 Print edition: ISSN 2188-1758 Paper New image processing pipelines for membrane detection Rajeswari Raju * Member, Tomas Maul † Non-member Andrzej Bargiela ‡ Non-member (Received December 8, 2014, revised January 5, 2015) Abstract: In this paper, we report an interesting observation about Denoising suggested by optimization ex- periments. Denoising is usually performed in order to minimize the detrimental effects that noise has on the subsequent stages of an algorithm. Thus Denoising is typically carried out as an early pre-processing stage before other core functions are applied. In the context of optimizing image processing chains for membrane detection, we gathered statistics of processing chains which exhibited an average F1 score larger than 90%, and observed that not one was found to use a Denoising function as its 1 st step in the processing chain. On the contrary, the optimization process tended to choose Denoising as a middle processing component, and generally selected im- age enhancement as an earlier component. We conclude, that at least in the context of this membrane detection problem, it is better to enhance information (enhancement) before cleaning it (filtering). Keywords: membrane detection, denoising, segmentation, image processing, optimization. 1. Introduction One of the aims of our research is to identify the best possi- ble sequence of image processing functions; capable of effi- ciently and accurately detecting neuronal membranes whilst ignoring and/or removing extraneous organelles from the processed output [1]. The problem of membrane detection, which can be seen to belong to the general class of segmen- tation problems, is characterized by several issues, includ- ing over and under segmentation due to similarities between membrane and non-membrane material. Many algorithms depend on ground-truth for training and require large num- bers of labelled training samples which is expensive and generally involves several time consuming processes [1]. In order to detect membranes whilst eliminating extraneous organelles we have proposed an approach called Image Pro- cessing Chain Optimization (IPCO). This approach (1) at- tains competitive accuracy levels whilst not requiring an ex- cessively long tuning phase, (2) does not require specialized hardware, (3) leads to chains consisting of short sequences of basic processing steps which are efficient and easy to in- terpret [1], (4) is simple to use [1] and (5) is flexible and can be applied to many different types of datasets. In carrying out our experiments we have discovered several interesting facts about optimal image processing chains, some of which * University of Nottingham Faculty of Science, School of Computer Sci- ence, Malaysia Campus, Semenyih, 43500 Selangor, Malaysia University Technology MARA, Faculty of Computer and Mathematical Science, Malaysia (khyx1rru@nottingham.edu.my) † University of Nottingham Faculty of Science, School of Computer Science, Malaysia Campus, Semenyih, 43500 Se- langor, Malaysia ‡ University of Nottingham, Faculty of Science, School of Computer Science, Jubilee Campus, Nottingham, NG81BB,United Kingdom are presented in this paper. When carrying out our research in membrane detection and organelle elimination, where activities ranged from manual fine-tuning [1] to automated segmentation using IPCO, we found that, at least for this membrane detection problem, Denoising typically appears later in the sequence (or chain) of processing functions. Moreover, in 10 cases out of 10 (including the best chain), we find contrast en- hancement before Denoising, suggesting that details need to be enhanced before cleaned, which could be encapsu- lated by the heuristic “enhance it before you lose it”. [2] In many cases even classification is done before Denoising. This paper is structured as follows: section 2 gives a brief overview of image processing workflows with special em- phasis on denoising; section 3 provides an overview of the IPCO approach and the data-set used; section 4 describes experimental results and preliminary analyses; section 5 concludes the paper. 2. Background Study According to Rafael C. Gonzales et.al [3], image analysis is a research area lying somewhere in between image pro- cessing and computer vision. According to this there are three types of processing, distinguished by different levels of abstraction, which are: low-level; mid-level; and high- level. Low-level processing involves: image preprocessing to reduce noise, contrast enhancement, and image sharpen- ing. Mid-level processing involves: image segmentation, description of objects in a form suitable for further compu- tational processing and classification or recognition of in- dividual objects. Finally, high-level processing involves: making sense of an ensemble of recognized objects and per- forming cognitive functions associated with human vision. In this paper, we focus on a crucial low-level component, Published by IIAE. 2015 15