Automatic Classification of Spider Images in Natural Background

Spiders are the most abundant predatory natural enemies in terrestrial ecosystems. As an important natural enemy of many agricultural and forestry pests, spiders play a very significant role in the biological control of pests. In order to make rational use of spider resources, it is necessary to observe and study the population characteristics of it. Direct observation method is time-consuming and laborious. If we can take the videos or images of spiders by surveillance cameras, and then use computer vision technology to identify and classify automatically, the efficiency of image data acquisition and pest biological control will be greatly improved. Motivated by this, we studied the classification and recognition of spider images in natural background obtained by common surveillance equipment. However, the images of some species of spiders in natural background are difficult to be collected, and the inadequate clarity and contrast of the subjects in images will also affect the recognition accuracy. So, firstly, histogram equalization was used to enhance the contrast of the image; the dataset of spider images was expanded by image rotation, reflection, flipping, zooming, translating and increasing the pixel noise appropriately, and so on; the contour detection was carried out for assistant recognition. Secondly, taken the deep convolutional neural networks (CNN) as our basic framework, two automatic recognition models of spider images, that is the 8-layer deep CNN model and the transfer learning model based on Inception-v3, were constructed. After that, the two models were trained, evaluated and compared under a dataset with 4478 pre-processed images. The experimental results show that the first model has a limited effect on image feature extraction of spiders in natural background, while the second model based on transfer learning can achieve better recognition accuracy when combining image contour features as an auxiliary input. In the second model, the accuracy of training set and testing set can reach more than 90%, and the recognition speed can be controlled within 1 second, which meets the practical requirements of automatic classification and recognition of spider images in natural background.