In the supervised learning object or image identification process, it is necessary to train a large number of images to perform the prediction or search. It is a well-known fact that arranging such a large amount of data for supervised learning is not always feasible, and this insufficient data often leads to poor training results. When the data is insufficient, image classification employs various clustering and classification models like K-means and K nearest neighbors to overcome this challenge, which eventually works on the distance based techniques oftenly yield unsatisfactory results. Hence, few-shot learning models play a vital role in developing a well-trained model using a moderate or low amount of data. The augmentation process always powers few-shot learning, generating different views of an image and increasing the number of implicit data points to dynamically boost the image identification process. Hence, the proposed model utilizes the two convolution neural networks to train the datasets. Initially, we deploy a 2D convolution neural network on a large dataset to obtain the trained model. Subsequently, we train the few-shotdataset with a channel-boost convolution neural network to enhance the channel features of the selected images. The obtained two models tend to hybridize using channel factors to provide the best matching images for the testing input images. This research results are analyzed with some factors to unleash the difference between augmented and unaugmented training. Here, the proposed model provides good results in the initial criteria thereby improving the accuracy in Few-shot image cleaning.