Article Preview
Top1. Introduction
Controlling the welfare of individuals, or a group of animal, is the main responsibility of an image-based animal supervision system (asup) (Khoramshahi et al., 2013). An asup, as an intelligent routine, seeks to provide an agricultural decision-support system (ADSS) with animal’s status information by analyzing diverse sources of inputs including image data. Image processing is gradually finding its rule in an asup as a result of decreasing in the processing cost. The state of the art image processing algorithms brings this opportunity to be more feasible and reliable than any other time. Many methods have been developed to use the power of image processing to measure the animal related parameters (such as the image-based movements) in order to give an estimation about a behavior coding (such as aggression, or farrowing).
Although the progress in the concepts of image-processing solutions, that are often acceptable for a controlled situation, is impressive, there is still a remarkable gap between controlled conditions, and real capturing settings that are under the effect of noise and many unwanted factors such as inconsistency in lightening condition or unwanted moving particle, which makes it a good target for developing methods that are more robust to noise, and less dependent on their training conditions.
Digital images, in a real capturing situation, are exposed to a variety of non-physical distortions during acquisition, such as the moving particles, or the poor-lighting condition, as well as the distortions during the processing, compression, storage, transmission, and the reproduction stages. Many of these factors can degrade the visual quality of an image (Wang et al., 2004).
The aim of a quality assessment (QA) research is to propose an algorithm to formulate the effective parameters in order to measure the quality of an image automatically and robustly (Sheikh et al., 2005), and take care or remove a LQ case from an image sequence.
According to the approaches that are employed to quantify the quality of an image as a numeric value, the QA methods are categorized into the following main groups:
- •
The full reference (FR) methods, which a set of reference images are processed to find the standards for checking the quality of an image.
- •
The non-reference (NR) methods, that try to find the quality measure with an unsupervised scheme (such as maximizing non Gaussianity by a tool such as kurtosis).
Between the parameters that affect the quality of a surveillance image, the effect of artifacts and the exposure change is covered in this article. The FR-QA methods for the effect of other factors, such as, sharpness, contrast, distortion, and exposure can be found for example in Sheikh et al. (2004), Saad et al.(2012), and Eckert et al.(1998).
The main physical factors, in an asup, that degrade the quality of an image can be considered according to the following categories:
- •
Changing and fluctuating in the lighting condition (e.g. changing a light switch).
- •
Small moving particles (e.g. mosquitoes, flies, and bugs).
- •
Unexpected movements in the borders. (e.g. opening/closing a pen's door).
- •
Unwanted objects inside the scene (e.g. human worker).
In this study, only the first three groups will be addressed, so the forth group is remained for further studies. The percentage of the frames, that are affected by the first three groups, are relatively low (in many experiments less than 5%), but even this small percentage negatively affects the accuracy and the precision of the consecutive processing routines (such as movement estimator), so within the preprocessing step of an image processing routine, they should be accurately recognized and filtered.