Conventional image capture approaches usually try the best to obtain high-resolution images to allow people to see details of the objects. Still, the high-resolution images do not always imply efficient classification accuracy. Holistic information of images has a more critical influence on the identification of the corresponding objects than detailed information of images. In this paper, we propose to attain optimal pixel re-representations of images and exploit them to perform classification. The merits of the proposed approach are as follows. (1) It automatically selects optimal re-representations for original images from a number of candidates by virtue of a feasible procedure. (2) Maximizing the ratio of the between-class distance to the within-class distance not only allows the optimal re-representations to be attained but also can adequately prompt sparse representation and collaborative representation to attain a better classification accuracy. The procedure of maximizing the ratio makes the obtained optimal re-representations of all objects as diverse as possible. Therefore, the attained linear combination of all training samples, in the form of vectors, of all objects can better express the test sample and has less expression error in comparison with the corresponding case based on original training samples of all objects. The experiments also prove that the optimal pixel re-representations are helpful to improve not only the sparse representation and collaborative representation but also other classification image classification approaches, to obtain more satisfactory classification accuracy.

常规的图像捕获方法通常尽最大努力获得高分辨率图像，以使人们能够看到物体的细节。尽管如此，高分辨率图像并不总是意味着有效的分类精度。图像的整体信息比图像的详细信息对识别相应对象的影响更大。在本文中，我们建议获得图像的最佳像素重表示，并利用它们进行分类。所提出的方法的优点如下。（1）通过可行的程序，它自动从许多候选对象中为原始图像选择最佳的重新表示。（2）最大化类间距离与类内距离的比率，不仅可以实现最佳的重新表示，而且可以充分提示稀疏表示和协作表示，以获得更好的分类精度。最大化比率的过程使所获得的所有对象的最佳重新表示尽可能地多样化。因此，与基于所有对象的原始训练样本的相应情况相比，所有对象以向量形式获得的所有训练样本的线性组合可以更好地表达测试样本，并且表达误差较小。