DCT and SVD Sparsity-Based Compressive Learning on Lettuces Classification

Abstract

Compressive Sensing (CS) technique in image compression represents efficient signal which offering solutions in image classification where the resources are constrained especially on a large image processing, storage resource, and computing performance. Compressive learning (CL) is a framework that integrates signal acquisition via compressed sensing (CS) and machine/deep learning for inference tasks directly on a small number of measurements, On the other hand, in real-world high-resolution (HR) data, where the image dataset is very limited CL, has the drawback of reduced accuracy under conditions of aggressive compression ratio. Here, a reconstruction method is necessary to maintain high levels of accuracy. To address this, we proposed a framework Deep Learning (DL) and Compressive Sensing that processing a small dataset of 92 images maintaining high accuracy. The framework developed in this paper employs processing sensing matrix A in compressive sensing with two transformation methods: DCT CL with Multi Neural Networks and the SVD method with GoogleNet framework. To maintain the same computation efficiency as DCT Compressive learning, SVD with GoogleNet framework provides a solution for object recognition, achieving accuracy values ranging from 89.47% to 63.15% for compression ratios of 3.97 to 31.75. This performance shows a linear tendency concerning the PSNR level, an index of signal reconstruction quality, and demonstrates an efficient process in the S matrix.

Keywords

• Compressive Learning
• Accuracy
• Discrete Cosine Transform
• Singular Value Decomposition
• Compressive Sensing