As a "paper exercise," weighing pallets is a reasonable approach. Assuming you know the individual case weight and non-case weight (e.g., pallet, cardboard, etc.), calculating number of cases from scale weight is straightforward.
In the "real world" you will have to deal with scale performance -- accuracy and repeatability -- over the range of gross pallet and case weights to verify if you can reliably resolve an individual package. Then there is problem of knowing the case weight -- will this be provided by the supplier, and be reliable, or will you have to take one off the pallet and weigh it? And then pallet and other packing variables. And calibration.
There is also installation of a platform scale. Will it be level with the floor? If so, it must be capable of the total weight that may be driven over it: maybe the full weight of a forktruck, depending on where and how it is installed. If raised, it is at risk to be hit by a pallet or other object.
In my experience, assuming you cannot get suppliers to provide assistance such as RFID tags, weighing pallets may be your least worst option. Machine vision will be tricky if having to deal with the variation shown in the original picture.
In the "real world" you will have to deal with scale performance -- accuracy and repeatability -- over the range of gross pallet and case weights to verify if you can reliably resolve an individual package. Then there is problem of knowing the case weight -- will this be provided by the supplier, and be reliable, or will you have to take one off the pallet and weigh it? And then pallet and other packing variables. And calibration.
There is also installation of a platform scale. Will it be level with the floor? If so, it must be capable of the total weight that may be driven over it: maybe the full weight of a forktruck, depending on where and how it is installed. If raised, it is at risk to be hit by a pallet or other object.
In my experience, assuming you cannot get suppliers to provide assistance such as RFID tags, weighing pallets may be your least worst option. Machine vision will be tricky if having to deal with the variation shown in the original picture.