Dear all,

Here is the $2023$ challenge. Have you heard about CAPTCHAs? Maybe not. But you have seen them all over the Internet for sure. Google CAPTCHA to learn more. Your task is to make a function called my$\_$classifier.m that decodes three digits in an image. You have plenty of training data in imagedata$/$ and in labels.txt you have the ground truth decoding.

images$\_$code$\_$lab$5.$zip Download images$\_$code$\_$lab$5.$zip

Your solution should be programmed reasonably well so it terminates within $5$ minutes on the training data set. We will run your function my$\_$classifier.m on our own data, which is similar but not identical to your training data. If you score above $40\%$ average precision and fulfill the time condition, then you pass the lab. See the enclosed script evaluate$\_$classifier.m and the example my$\_$classifier.m for an explanation of the expected formats. You are not allowed to use the built$-$in OCR function in Matlab in your solution. You can use the Deep Learning toolbox. You are allowed to use only Matlab to make the comparison fair.

Zipfile description:

Training data in $/$imagedata

Labels in labels.txt

Matlab files: evaluate$\_$classifier.m$,$ my$\_$classifier.m

Guidelines:

Leave evaluate$\_$classifier.m unchanged except for the parameters.

Train your classifier with the training data in $/$imagedata in your own file, for example, train.m$.$

Save the weights as a Matlab data file, named, for example, model.mat.

Load the weights for the classifier to run, that is$,$ modify my$\_$classifier.m$,$ however, do not change the function call in my$\_$classifier.m$.$

Submit train.m$,$ model.mat and my$\_$classifier.m as a zip file in Studium.

You should be able to run the evaluate$\_$classifier.m on the training data. Training data can only be used during the training time and should not be used in any way when running evaluate$\_$classifier.m$.$ We will not train your model, only evaluate it$,$ which is why you have to submit a working classifier. Please, know that we have our own test data to test your classifier.

Upload your solution $($the $.$m$-$file$($s$))$ to Studium along with your own cool team name. You should put the estimate of the average precision for the training data and the execution time of the method for training data as a comment in your submission in Studium. No report is required!