Defect inspection is one of the most important processes for the touch screen manufacturing. Because the affections of uneven illumination, camera resolution, defect types and the textural background of the touch screen images, the defect inspection problem of touch screen becomes complex and the accuracy of inspection rate is effected significantly. In this paper, the features from defect-free touch screen images are collected to generate an atom pool. Since the normal feature pool is redundant, an optimal subset with small size is selected from the atom pool as training dictionary. According to the l1 minimization, the coefficients for sparse linear representation of a testing image under the redundancy dictionary can be obtained. Thus, the defect inspection problem can be transferred to the problem that if an image can be sparsely represented under the redundancy dictionary or not. Sparsity ratio of the sparse representation coefficients is proposed as a measurement to determinate whether the testing image is defective or not. Experimental results show that under various illumination conditions, the proposed approach can efficiently and quickly detect the touch screen defects for low resolution images and different defect types.