LP Constrained Softmax Loss for Handwritten Image Recognition

Handwritten digit image recognition is a task that integrates two domains- computer vision with sequence learning. For the past recent years, the performance of handwritten image recognition or rather face verification has improved using the convolutional neural network (CNNs). In a typical handwritten digit image recognition method, normalization of features is an expository measure for enhancing the performance. This prompts us to instigate the outcomes of normalizing the features in the training phase. We propose a strategy for feature normalisation that is a modification of the existing softmax loss by adding a p^th order constraint that we call Lp constraint that optimizes the normalised variant of inner product called cosine similarity. This constraint has remarkably increased the performance of the softmax loss. The accuracy is increased by 6.27% with respect to the L2 constrained softmax loss and largely increased w.r.t the regular softmax loss. This significant step boosts the performance of handwritten digit image recognition.