IMAGE PREPROCESSING Clause Samples
The 'Image preprocessing' clause defines the procedures and requirements for preparing images before they are used in a particular process or system. This typically involves steps such as resizing, normalization, noise reduction, or format conversion to ensure that images meet specific technical standards or are compatible with downstream applications. By standardizing how images are handled prior to use, this clause helps ensure consistency, improves the quality of results, and reduces errors caused by incompatible or poor-quality image data.
IMAGE PREPROCESSING. Preprocessing plays a large role in determining the classification results obtained within this project when using appearance based techniques for representing or matching images. This is mainly due to the outdoor, far field nature of the AVITRACK project (which adds large amounts of perspective distortion and lighting variations), as well as the low contrast pictures obtained from the cameras used. Image pre processing allows the effects of perspective distortion to be minimised using scaling algorithms, as well as limiting the effects of lighting variations. Preprocessing has a computational cost attached to it; however, if a suitable balance can be found between cost and benefits, final recognition rates can be improved without unduly reducing the overall speed. In this project, different rescaling techniques were trialled, with differing computational costs associated with them, from the very low (nearest neighbour interpolation), to the highest (bicubic interpolation). Although the cost per use for each interpolation method is quite low overall (as compared to, for example, locating scale invariants), the images within a matching system often requires multiple rescaling before a classification may be decided upon. This therefore increases the time required by a factor equal to the number of images to resize, and which grows with increased image size. Within this project, it was found experimentally (See [85]), that little difference was evident between bilinear and bicubic rescaling methods, both in visual examination and classification rates. It was therefore decided that the extra smoothing available within the bicubic scheme was superfluous to requirements and that to maximise efficiency and classification rates, bilinear interpolation may safely be used. Having completed work using Gaussian scale space, it is possible that an alternative to using bilinear or bi cubic scaling would have been to use nearest neighbour interpolation then smoothed the image using a Gaussian convolution. This idea is normally referred to as a Gaussian pyramid. The drawbacks to using this type of technique are that it is most efficient when the desired size is a power of two from the original image. This is unlikely to occur within the AVITRACK image files, as the images are often nonsquare and are of arbitrary size. It should also be noted that the blurring achieved by the Gaussians within a Gaussian pyramid is recreated using pixel averages within the bilin...