A Lightweight Underwater Object Detection Model: FL-YOLOV3-TINY

Tags

Top Highlights

• deep separable convolution

• deep separable convolution

• This paper replaces the main feature extraction module of YOLOV3- Tiny network with the model Bottleneck[23] designed based on Deep Deposable Convolution technology.

• the dimension of the whole network layer is low, the computation speed will be accelerated and the performance of the whole model will be better

• However, it is difficult to obtain sufficient information by using only low-dimensional tensor to extract features. Therefore, we hope that enough information can be extracted to achieve good results. Bottleneck will balance this problem, before the deep separable convolution, the dimensions are promoted through the Expansion layer to extract features from the model. After that, Proj ection Layer is used to compress and reduce dimensions to make the network model more lightweight.

• he Bottleneck module replaces the traditional convolution and pooling layer, and the detection layer adds a 54*54 detection scale compared with YOLOV3- Tiny, which can extract small object image features more fin