Convolutional Neural Networks (CNNs) are being used popularly for detecting and classifying objects. Rotational invariance is not guaranteed by many of the existing CNN architectures. Many attempts have been made to acquire rotational invariance in CNNs. Our approach ‘Eigenvector Orientation Corrected LeNet (EOCL)’ presents a simple method to make ordinary LeNet [1] capable of detecting rotated digits, and also to predict the relative angle of orientation of digits with unknown orientation. The proposed method does not demand any modification in the existing LeNet architecture, and requires training with digits having only single orientation. EOCL incorporates an ‘orientation estimation and correction’ step prior to the testing phase. Using Principal Component Analysis, we find the maximum spread direction (Principal Component) of each test sample and then align it vertically. We demonstrate the improvement in classification accuracy and reduction in test time achieved by our approach, on rotated-MNIST [2] and MNIST_rot_12k test datasets, compared to other existing methods. © 2020, Springer Nature Singapore Pte Ltd.