Design of a brief perceptual loss function with Hadamard codes

Perceptual loss functions are central to an ever-increasing number of tasks across computer vision. Their strength lies in their ability to translate perceptual nuances into numerical high-level features. A cornerstone of these functions are the high-dimensional, real-valued deep feature vectors. However, their memory-intensive nature often hinders deployment on devices with constrained resources. We introduce a concise perceptual loss function underpinned by Hadamard codes. For the ImageNet collection, our method delivers a lean representation of a mere 128 bytes. Impressively, this representation is not tied to any specific architecture, paving the way for the integration of industry-standard models. Utilizing our proposed binary codes in conjunction with k NN and Half-Space Proximal (HSP) classifiers (with HSP being a noteworthy alternative to k NN), we have secured commendable accuracy. This novel approach sets new benchmarks, enhancing state-of-the-art performance in knowledge transfer across a variety of image datasets.