step1: make a standardized template library (all fruit, 51x51).
step2: extract an individual object from an image and generate a single line edge (contour).
step3: standardize the image (normalized, single line, 51x51).
step4: use the central point as fix point, clockwise scan the two template images (contour projection).
step5: choose a tolerance value (3 or 5 pixels) to evaluate the image with each template, and get a score (contour matching).
step6: decide what kind of fruit it is by lowest score.
i can take the template as a kernel and by applying it to the different fruits using 'corr' OR 'conv'or something else , which i can get to know after more details about the exercise, then after which we can hold a threshold for the fruit detection.