Wood biomass is a crucial resource in the field of bioenergy. It is commonly used as a feedstock in bioenergy generation and biochemical production. However, to fully make use of the characteristics of distinct woody biomass, before delivering it to subsequent processing steps, it is necessary to classify and distinguish the different classes of wood biomass. Currently, companies tend to hire operators to carry out the classification task with the help of a digital inventory of biomass classes. However, operators require a great deal of training to correctly group the wood biomass into broad groups with relatively similar quality. Moreover, the judgment bias of operators and the small differences among materials, such as the visual similarity of shavings and sawdust, can lead to potential classification inaccuracy. Considering these challenges, it is inevitable to explore an alternative solution. Thanks to the low cost of the collection of digital images, this work proposes the use of deep neural networks to automatically and precisely classify images. However, this is a challenging task. Specifically, the discrimination in features among certain classes is small, especially with the involvement of some unexpected backgrounds. To address the challenge, compared with the traditional practice that a mere feature vector is extracted for an image, a feature extraction module is applied to produce sets of diverse feature vectors from a single image to explore more visual information. Self-attention mechanisms are intensively utilized in the stage of extraction. The effectiveness of the proposed method is experimentally shown in our collected data set. After a voting strategy is applied, the average accuracy of our method is 96.2%.