Wild animal recognition based on effective-class-balanced softmax loss

Abstract

Wild animal recognition is important for wild animal protection. Because the number of different wild animals is different in the wild. The wild animal image dataset collected in field by using camera trap is a typical long tail dataset. This paper proposes an Effective-Class-Balanced Softmax Loss (ECBSL) to solve the long tail problem of self-built wild animal dataset. Firstly, a new cross entropy loss function is obtained by using pointwise mutual information instead of conditional probability for modeling. Then the improved effective number of samples calculation method is used to approximately calculate the prior probability distribution of different animal species. Finally, the effectiveness of ECBSL is proved by experiments. Experiments on the self-built wild animal dataset show that the proposed method improves the recognition accuracy of the tail classes and the whole dataset. The comparison experiments with other methods show that the proposed method is superior to other methods.