Although deep learning has achieved remarkable successes over the past years, few reports have been published about applying deep neural networks to Wireless Sensor Networks (WSNs) for image targets recognition where data, energy, computation resources are limited. In this work, a Cost-Effective Domain Generalization (CEDG) algorithm has been proposed to train an efficient network with minimum labor requirements. CEDG transfers networks from a publicly available source domain to an application-specific target domain through an automatically allocated synthetic domain. The target domain is isolated from parameters tuning and used for model selection and testing only. The target domain is significantly different from the source domain because it has new target categories and is consisted of low-quality images that are out of focus, low in resolution, low in illumination, low in photographing angle. The trained network has about 7 M (ResNet-20 is about 41 M) multiplications per prediction that is small enough to allow a digital signal processor chip to do real-time recognitions in our WSN. The category-level averaged error on the unseen and unbalanced target domain has been decreased by 41.12%.
Wireless sensor networks (WSNs) typically are designed to detect and identify neighboring objects in wild [1, 2, 3, 4] with sound or vibration sensors in the form of single  or microarrays . The sound or vibration sensor has many advantages [7, 8, 5], such as low cost, low energy consumption, and relatively low in algorithm complexity. However, they are unsuitable for mixed objects detection because their spatial resolutions are usually too low to distinguish each person in a group of pedestrians. To overcome this shortage, we have employed cameras in our WSNs which has been proved to be effective for dense targets identification . Unfortunately, images captured by WSNs are noisy, such as low in illumination, resolution and photographing angle, which are different from most publicly available datasets. Because the severe limitation in data and resources, despite the rapid development in deep learning [10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], WSN-applicable deep-learning-based image classification algorithms evolve slowly. So, cost-effective dataset construction methods are needed urgently to build datasets that corresponding to specific WSN applications. Several random images of the target application (target domain) that was captured during our field experiments have been shown in Fig. 1, where targets like persons and cars are hard to identify.
Because of limited communication bandwidth, WSNs cannot run deep neural networks (DNNs) in a remote cloud (or fog) which is a common strategy for embedded devices [21, 22, 23, 24, 25]. To run DNNs in such devices locally [26, 27], a training strategy is wanted to cut computation costs without decreasing identification accuracy significantly. Fortunately, Han et al. [28, 29] have pointed out that only parts of weight parameters in neural networks are playing essential roles during predictions. Therefore, it is possible to train an efficient DNN for WSNs with fewer parameters if we can fully utilize key weight parameters.