Improvements Based on Feature Fusion Single Shot Multibox Detector

Abstract

As people have paid more and more attention to deep learning, the development of deep learning is also faster and faster. However, when the object is detected, the accuracy and real-time of the object result is still a challenging problem. For the problem of scale variation in object detection, we mainly use the method of feature pyramid to solve this problem, which has been applied in many detectors. At present, there is an object detection algorithm which is very high speed and high precision. FSSD (Fusion Single Shot Multibox Detector) algorithm has an obvious disadvantage that it cannot obtain more accurate positioning information for medium-sized objects and lower-level information. Especially after feature fusion, the semantic information expression in the shallow layer of its backbone network is not complete enough. This paper proposes a combination of bottom-up pathway, fusing-splitting pathway and RFB (Recepive Field Block) module, through the combination of the above modules, it is actually a relatively effective FSSD model architecture. RFB module can expand receptive field and improve target detection accuracy. Furthermore, in the whole FSSD model network, the architecture is built by merging the split path and RFB module through the bottom-up path. Two different approaches of the multi-scale feature pyramid introduce both better positioning tips and more medium-sized objects information. RFB module introduce finer high-level semantics on shallow layers. On the primary detector FSSD, we used the above method to carry out placement detection. The results show that our method is feasible and the result is the most advanced.