Data processing for image-based chemical sensors: unsupervised region of interest selection and background noise compensation.

Natural olfaction suggests that numerous replicas of small sensors can achieve large sensitivity. This concept of sensor redundancy can be exploited by use of optical chemical sensors whose use of image sensors enables the simultaneous measurement of several spatially distributed indicators. Digital image sensors split the framed scene into hundreds of thousands of pixels each corresponding to a portion of the sensing layer. The signal from each pixel can be regarded as an independent sensor, which leads to a highly redundant sensor array. Such redundancy can eventually be exploited to increase the signal-to-noise ratio. In this paper we report an algorithm for reduction of the noise of pixel signals. For this purpose, the algorithm processes the output of groups of pixels whose signals share the same time behavior, as is the case for signals related to the same indicator. To define these groups of pixels, unsupervised clustering, based on classification of the indicator colors, is proposed here. This approach to signal processing is tested in experiments on the chemical sensitivity of replicas of eight indicators spotted on to a plastic substrate. Results show that the groups of pixels can be defined independently of the geometrical arrangement of the sensing spots, and substantial improvement of the signal-to-noise ratio is obtained, enabling the detection of volatile compounds at any location on the distributed sensing layer.