Image recognition has experienced enormous progress in research and development over last two decades and it has its applications in many fields like security, medical, banking, automation etc. The existing method generally used for image recognition is the template matching, where in the image is recognized by comparing it with the already stored images in the database. But the main drawbacks with this method are when the image is affected by noise, it cannot be recognized and more bandwidth is required, since the complete image has to be sent. Large storage capacity is needed when more number of images is to be recognized. These drawbacks can be overcome by using artificial neural network where in recognition can be done by using only certain parameters of the image. Even this method fails when the image undergoes scaling, rotation or translation. Scaling can be done by normalization, but translation and rotation (above 3degress) cannot be done by normalization, so statistical moments are used where in invariant moments are trained by using artificial neural networks. Here the two methods, training of artificial neural networks with certain parameters and training of invariant moments using artificial neural networks are compared for image recognition against translation, rotation, and scaling. The results show that the proposed method gives the better performance.