Compact CNN Model for Classifying Rose Apple Species and Detecting Their Skin Defects

For deploying a classified application on a machine system, the system should accurately recognize, fast compute, and easily deploy on hardware. However, most of the convolutional neural network (CNN) architectures that can perform a high performance are so complex and have very deep layers. Since these models have a large number of parameters and slowly compute therefore they are difficult to deploy on hardware that doesn't have a high-performance processor. This paper presented the compact CNN model for classifying rose apple species and also detecting their skin defects as well as the traditional deep CNN models such as ResNet50 and DenseNet121. Comparing our model with ResNet50, our model reduced the number of parameters and prediction time from ResNet50 by 99.38% and 80.91% respectively. Comparing our model with DenseNet121, our model also reduced the number of parameters and prediction time from DenseNet121 by 97.93% and 81.81% respectively. The overall accuracy of our model was 97.90% that nearly to ResNet50 and DenseNet121, it only less than 0.34% and 0.72% respectively. From these results, they were strong evidence that the compact CNN model not only had a high performance as well as the traditional deep CNN models but also had a small number of parameters and fast computed.