A Novel Hyperparameter Optimized Quantized MobileNet Architecture for Improving Diagnostic Accuracy in the Medical Image Classification

In recent years, there is a remarkable use of smartphones in the field of computer vision. Deep convolutional neural networks being the backbone of computer vision has forced the DCNN to mitigate from heavyweight network to lightweight network. Smartphones are resource constrained devices with limited storage and processing capabilities. DCNN, a heavyweight network, cannot be implemented directly on such devices. We require a lightweight version for smartphones that meets the desirable properties such as small model size and acceptable accuracy. DCNN can be compressed using the neural network quantization process. But the quantization process introduces quantization error that degrades the accuracy. In this paper we propose a novel hyperparameter optimization of quantized deep convolutional architecture that mitigates quantization loss while maintaining the competing accuracy. The performance of HPOQ-MobileNet is tested and verified for image classification tasks on Diabetic Retinopathy dataset. The accuracy of the baseline model is 83.81% and that of proposed HPOQ-MobileNetv2 architecture is 82.9%. The inference accuracy of fine-tuned Mobilenetv1 is 81.8% and that of proposed HPOQ-MobileNetv1 model is 80.3%. The proposed architecture shows optimized accuracy and effective resource utilization by reducing the memory footprint by 4×, and quantization loss approximately by 1%. Quantization simulation shows that the model can be ported to resource constrained devices.