Task-specific CNN size reduction through content-specific pruning

Nurbek Konyrbaev^1*Martin Lukac²Askhat Diveev³Elena Sofronova³Sabit Ibadulla¹Asem Galymzhankyzy¹

• ¹Korkyt Ata Kyzylorda State University, Kyzylorda, Kazakhstan
• ²Hiroshima City University, Hiroshima, Japan
• ³Russian Academy of Sciences (RAS), Moscow, Moscow Oblast, Russia

The widespread and growing use of flying unmanned autonomous vehicles (UAV) is attributed to their high spatial mobility, autonomous control, and lower cost when compared to usual manned flying vehicles. Applications such as surveying, searching, or scanning the environment with application-specific sensors have made extensive use of UAVs in fields like agriculture, geography, forestry, or biology. However, due to the large number of applications and types of UAVs, limited power resources have to be taken into account when designing task-specific software for a target UAV. The power constraints of, in particular, smaller UAVs will in general hinder their mobility, limit the sensors, or limit their distance from the control transmitter. These constraints will in general require that the UAV have a high level of autonomy. Reducing the overhead of the control and decision-making software onboard is one approach to increasing the autonomy of UAVs. Specifically, we can optimize and minimize the task-specific design of onboard control software more effectively than a general-purpose algorithm. In this work, we focus on reducing the size of the computer vision object classification algorithm. We define different tasks by specifying which objects the UAV must recognize, and we construct a convolutional neural network (CNN) for such a specific classification. However, rather than creating a custom CNN that requires its dataset, we begin with a pre-trained general-purpose classifier. We then choose specific groups of objects to recognize, and by using response-based pruning (RBP), we simplify the general-purpose CNN to fit our specific needs. We evaluate the pruned models in various scenarios. The results indicate that the evaluated task-specific pruning can reduce the size of the neural model and increase the accuracy of the classification tasks. For small UAVs intended for tasks with reduced visual content, the proposed method solves both the size reduction and individual model training problems.