PROJECT TITLE :

GPCA A Probabilistic Framework for Gaussian Process Embedded Channel Attention

ABSTRACT:

It is common practice to employ channel attention mechanisms in a variety of visual tasks in order to achieve effective performance improvements. It is able to bolster the channels that provide informative content while simultaneously suppressing the channels that provide no value. In recent times, numerous channel attention modules have been suggested, each of which can be implemented in a distinct manner. In a general sense, they are predominately founded on the operations of convolution and pooling. In this paper, we propose the Gaussian process embedded channel attention module, also known as the GPCA module, and further interpret the channel attention schemes in a probabilistic manner. The GPCA module's goal is to model the correlations between the channels, which are presumed to be represented by variables with a beta distribution. Because the beta distribution cannot be incorporated into the end-to-end training of convolutional neural networks (CNNs) with a solution that is mathematically tractable, we use an approximation of the beta distribution to solve this problem. This allows us to train CNNs more effectively. In order to be more specific, we make use of a Sigmoid-Gaussian approximation, in which the variables that are distributed according to a Gaussian are moved into the interval [0,1]. After that, the Gaussian process is applied to the task of modeling the correlations between the various channels. In this particular instance, a solution that can be solved mathematically is derived. The GPCA module can be effectively implemented and incorporated into the all-encompassing training of the CNNs in a seamless manner. The results of the experiments show that the proposed GPCA module has a performance that has a lot of potential.