ARC-CNN: An Accurate and Efficient Deep Learning Model for Apple Ripeness Using SE Modules and Depthwise Convolutions

Currently CNN models play a crucial role in image recognition particularly in agricultural applications. However balancing recognition accuracy and computational cost remains a challenge. To address this the paper introduces ARC-CNN an innovative CNN model based on ResNet-50 tailored for fruit ripeness classification. ARC-CNN leverages the residual connections of ResNet-50 to effectively mitigate gradient vanishing enabling the construction of deeper network structures. Additionally, it integrates two key innovations: the Squeeze-and-Excitation (SE) module and depthwise separable convolution. The SE module recalibrates feature channel interactions enhancing the model's ability to capture informative features. Meanwhile depthwise separable convolution significantly reduces computational costs and model size making ARC-CNN suitable for deployment on devices with limited resources. Our experiments demonstrate that in the task of apple ripeness classification ARC-CNN has achieved a significant improvement in accuracy compared to the standard ResNet-50 model. Specifically, the accuracy on the training set has been enhanced from 96% to 98% marking a 2% increase that underscores the model's enhanced capability to capture informative features. Notably ARC-CNN's floating-point operations per second (GFLOPS) have increased from 4.132 to 7.874 indicating a substantial boost in processing efficiency while maintaining computational costs closely comparable to those of ResNet-50.