Figure 2.
The architecture takes an aircraft input and sends it through three feature-extraction branches. The first branch uses a residual network. Its upper stages include convolution, batch normalisation, rectified linear unit and maximum pooling in Stage 1, followed by convolution and identity blocks in Stage 2. Stage 3 contains convolution and identity blocks. The first two stages are marked as frozen, while Stage 3 is marked as trainable. The second branch uses Efficient Net version 2. Its frozen section begins with a 3 by 3 convolution and continues through fused mobile bottleneck convolution and mobile bottleneck convolution blocks with several repeated stages and increasing channel dimensions. Later mobile bottleneck convolution blocks and the final 1 by 1 convolution, pooling and fully connected stage are marked as trainable. The third branch uses a Contrastive Language-Image Pretraining visual encoder followed by a visual encoder. Its output is arranged as a matrix of visual tokens labelled V 1 to V N and V 1 T 1 to V 1 T N, V 2 T 1 to V 2 T N, and corresponding later rows. This visual encoder section is marked as frozen.Proposed shared multi-branch network includes ResNet, EfficientNetV2 and CLIP visual encoder