Based on the above explanations, we can see that food recognition and calorie computation are essential components of EHSH. However, due to the resource constraints in the mobile device, implementing them in a smartphone was not viable option, since
the response time was very high, and the application consumed a large percentage of the smartphone’s resources, limiting the usability of the application and the smartphone. To overcome these challenges, we used the cloud to offload a bulk of the processing
tasks. In [6,7], we used MapReduce as a parallel classifier in the cloud to implement SVM on each of the cloud instances. As part of the SVM implementation in Amazon Web Services (AWS), we performed the map and reduce based training and testing tasks on
multiple EC2 instances. Although the MapReduce model yielded the desired results for larger training sets, the initial time taken to distribute the data and process it across different Amazon EC2 instances was considerably higher for smaller training sets of data. Based on the results in [5], we observed that with the increase in cloud instances (from a t2.micro one cloud instance to x3.large four cloud instances), the overall response time got reduced. In a scenario where higher number of users is accessing EHSH in parallel, adding more cloud instances on demand would suffice the demand for cloud resources, but the initial time taken to allocate the cloud instances would adversely affect the overall response time. Hence the scalability aspect was still a challenge that had to be addressed.