I fine-tuned an Inception V3 model provided in AWS SageMaker to detect COVID-19 Rapid Test Results (see the image below for an example). I provided about 20 pictures of negative and about 20 pictures of positive tests for the training. All pictures were taken with slightly changing angles and positions. However, when testing the fined-tuned model, the recognition did not work at all.
Is the deviation between both image classes is too small (only changing red bars). Is there any other way a machine learning model can detect such small deviations between a positive and negative test without heavy image pre-processing?