Algorithm1: Unsupervised trademark retrieval method based on attention mechanism

Input: Retrieved image I, Trademark database M. Output: Image sequence R which is similar to I. Step1: for i←1 to maximum_epochs do       1.  $\mathrm{Select} \mathrm{training} \mathrm{samples} \mathrm{from} \mathrm{M} \mathrm{to} \mathrm{obtain} \mathrm{the} \mathrm{training} \mathrm{batches} \mathrm{as} \mathrm{X}=\left\{{\mathrm{X}}_{\mathrm{i}}\right\},\mathrm{i}=1,2\cdots \mathrm{n}$.       2. $\mathrm{Feature} \mathrm{extraction} \mathrm{obtains} {\mathrm{v}}_{\mathrm{i}}$$\mathrm{to} \mathrm{form} \mathrm{feature} \mathrm{set} \mathrm{V}=\left\{{\mathrm{v}}_{\mathrm{i}}\right\}$, put V into the instance discrimination module.       3.  $\mathrm{Calculate} \mathrm{the} \mathrm{loss} \mathrm{from} {\mathrm{v}}_{\mathrm{i}}$ and optimize loss, update V iteratively.       4. Backpropagate the loss and update the parameters.       5. Repeat the above steps until the algorithm converges to get the feature extraction network N.    end for step2:       1. $\mathrm{Put} \mathrm{M} \mathrm{into} \mathrm{the} \mathrm{N}, \mathrm{form} \mathrm{a} \mathrm{database} \mathrm{of} \mathrm{trademark} \mathrm{image} \mathrm{feature} \mathrm{as} {\mathrm{F}}^{\prime }=\left\{{\mathrm{f}}_1^{\prime }{,\mathrm{f}}_2^{\prime }\cdots {\mathrm{f}}_{\mathrm{N}}^{\prime }\right\}{,\mathrm{F}}^{\prime }\in {\mathrm{R}}^{\mathrm{N}\times 128}$$, \mathrm{store} {\mathrm{F}}^{\prime }$ in the retrieval module.       2. $\mathrm{Put} \mathrm{I} \mathrm{into} \mathrm{the} \mathrm{N}, \mathrm{get} \mathrm{a} \mathrm{image} \mathrm{feature} \mathrm{as} {\mathrm{f}}_0^{\prime }\in {\mathrm{R}}^{1\times 128}$$, \mathrm{store} {\mathrm{f}}_0^{\prime }$ in the retrieval module.       3.  $\mathrm{Measure} \mathrm{similarity} \mathrm{between} {\mathrm{F}}^{\prime }$$\mathrm{and} {\mathrm{f}}_0^{\prime }$, output similar image sequence R.